Machine Learning and Knowledge Extraction doi: 10.3390/make8100333

Authors: Fernando Salazar Aryan Baibordy José Manuel González Juan Mata

Structural dam safety control is fundamentally based on visual inspections and on the analysis of monitoring data and their comparison with the expected behavior for a given load combination, which in turn is calculated with a predictive model. In recent decades, the development of machine learning algorithms has led engineers and researchers worldwide to explore their capabilities for predicting dam behavior and supporting safety assessment. Regardless of the model used, a prediction is generally obtained together with an expected model error, which is compared with the observed value in order to support decision-making. This process, however, requires the intervention of experts in dam engineering, and particularly in the prior behavior of the structure under analysis. In recent years, the emergence of generative artificial intelligence models has opened up considerable opportunities to improve processes in various fields, including civil engineering. The reported capabilities of these models make them well suited to some tasks in the field of dam safety, particularly for developing predictive models and interpreting their results. In this work, we tested ChatGPT to assess its potential in these tasks, using the data from the ICOLD 2022 Benchmark Workshop, in which 18 teams provided their best predictions of the behavior of an arch dam. ChatGPT was provided with the same problem description and monitoring data available to the participants and was asked, through brief and neutral prompts, to propose a modeling strategy, generate predictions for six response variables for short- and long-term periods, create the corresponding source code, and interpret the results from a dam safety perspective. The complete procedure was performed twice using the same information and prompts. The results show considerable potential for increasing productivity in building and applying machine learning models, with predictive accuracy comparable to that obtained by the Benchmark participants. ChatGPT achieved a lower mean absolute error than the participants’ average in 10 (first run) and 12 (second run) of the 12 variable–period combinations, and outperformed the best participant in one case. However, the interpretation of the results was considerably weaker, with important aspects being only partially addressed or overlooked. These findings highlight the potential of ChatGPT to assist in the development of data-based predictive models, but also reveal its limitations in terms of the domain-specific knowledge required to draw conclusions on dam safety.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100331

Authors: Jhony Fernando Lopez Torres Juan Gabriel Castañeda Edwin Rivas

Nutritional estimation pipelines built on large language models are entering biomarker research, but free-text meal descriptions have no ground truth, so outputs cannot be validated against a reference standard. What can be established without one is whether the same input returns the same number. We submitted 191 de-identified Spanish meal descriptions, byte-identically, three times each, at temperature 0, to 20 model configurations across two inference platforms (12,119 calls), and measured test–retest reliability by ICC(1,1), cross-model agreement, internal Atwater coherence between reported energy and macronutrients, and the reliability gradient across input specificity. All 20 models exceeded the audited cascade’s energy ICC under both comparators, but reproducibility did not imply internal coherence: the most reproducible model (ICC 1.0000) departed from its own macronutrients by more than 10% under the simplified 4/4/9 Atwater approximation in 65.58% of responses, against 3.80% for the cascade. The two axes are close to orthogonal, and the dissociation needs no threshold to state: that configuration ranks first of 20 in reliability and first of 20 in incoherence at every bound between 5% and 20%. A binary acceptance gate is not robust to that bound—3 to 17 of 20 configurations pass across the range—whereas the ordering is (Kendall tau ≥ 0.74 between any two), so coherence is reported as a ranked screening dimension with a decision curve, not as a validated criterion. Test–retest reliability is therefore necessary but not sufficient to authorise an LLM-assisted nutritional pipeline, and acceptance requires at least one reference standard-free coherence check alongside the reliability estimate.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100332

Authors: Galitsky

Large language models can produce convincing explanations that rely on invented facts, unsupported assumptions, or overlooked alternatives. This paper presents a method for checking whether such explanations are supported by the available evidence. The method identifies the assumptions needed to support a model's answer, checks for contradictions, and compares the answer with alternative explanations that require fewer unsupported assumptions. It also considers how an explanation is organized, distinguishing its central claims from supporting details and examining how statements support or challenge one another. Evaluation on four datasets shows that incorporating this structure improves the detection of hallucinated reasoning compared with the same detection approach without structural information. Additional experiments examine the contributions of explanation structure, assumption checking, and comparison with alternatives. The findings suggest that checking both the evidence behind an answer and the organization of its explanation can help detect and explain failures in language-model reasoning.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100330

Authors: Aleksei Samarin Artem Nazarenko Dmitry Nazarenko Julia Afonina Egor Kotenko Aleksandra Dozortseva Anastasia Mamaeva Alexander Savelev Galiya Narova Elena Mikhailova Alexander Motyko Valentin Malykh

Automated microscopic image analysis is essential for quantitative microbiology, yet manual acquisition and generic recognition models struggle with morphological variability and scene artifacts. We address both ends of this pipeline. First, we present a low-cost, open-source acquisition system, based on an Arduino UNO R3 microcontroller and NEMA 17 stepper motors, that automates focusing, magnification, and illumination on microscopes from the Levenhuk MED series, including the Levenhuk MED 20T, MED 25T, MED 30T, MED 35T, and MED 40T, and applies a standardized, programmable control procedure intended to improve the reproducibility of imaging conditions. Using this system, we construct a benchmark of 100,000 microscopy images across five classes (micrococci, diplococci, streptococci, bacilli, and background), with annotations for classification, detection, and segmentation. Second, we propose Microscopy-Aware Attention (MAA), which augments standard self-attention with six learnable, biologically motivated priors capturing cell boundaries, focus quality, internal-structure contrast, spatial compactness, shape, and local density. MAA is evaluated on six classification backbones, six detection backbones, and five segmentation backbones against no-attention, Squeeze-and-Excitation, spatial-attention, and CBAM baselines. It consistently yields the best results, raising weighted F1 for classification from 0.947 to 0.975 (DINOv3), mAP50 for detection from 0.901 to 0.928 (InternImage-H), and mIoU for segmentation from 0.834 to 0.862 (SegMAN), with component-wise ablations showing task-specific contributions of each prior.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100329

Authors: Tatiana Lazovskaya Dmitriy Tarkhov Igor Velikorechanin Galina Malykhina

Physics-informed neural networks usually incorporate differential equations primarily through the loss function, while the neural architecture itself is selected independently of the underlying numerical structure. This work develops a scheme-induced approach in which an analytically modified numerical method defines the computational architecture, while selected elements of the numerical scheme are made trainable and refined using physics-informed loss functions. The specific contribution is the formulation and systematic investigation of a class of scheme-parameterized low-fidelity physics-informed neural networks (LF-PINNs) for evolutionary PDEs. In the implementation studied here, a compact parameter network directly controls the coefficient of a generalized θ-scheme, while the implicit update is represented by a fixed number of Picard iterations. The approach is evaluated on six one-dimensional PDE benchmark configurations against conventional, hard-constrained, and adaptively weighted PINNs. The nine-parameter LF-PINN achieves competitive accuracy on several benchmark problems, while its trainable parameter remains explicitly connected to a coefficient of the underlying numerical scheme. The results also show that performance is problem-dependent and that a small number of trainable parameters does not necessarily imply lower computational cost. Stability, ablation, computational-cost, and generalization analyses further characterize the properties and limitations of the scheme-induced construction.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100328

Authors: Aldo Spolaore Jonas Scholl Athanasios Gkampenis Kathrin Machetanz Marcos Tatagiba Georgios Naros

Retrospective clinical research depends on reliable extraction of structured variables from unstructured medical documentation, but manual chart abstraction is slow and vulnerable to inter-rater variability. We compared zero-shot extraction of clinically relevant variables from German neurosurgical records by large language models (LLMs) with extraction by two human auditors. We analyzed 881 de-identified clinical notes from 211 patients undergoing microvascular decompression and extracted 15 predefined binary variables across three domains: preoperative symptoms, postoperative complications, and disease course. Six LLMs were assessed, including cloud-based models (GPT-4o, o3-mini, Claude Sonnet 3.5, and Gemini 2.0 Flash) and locally hosted models (llama 3.3 and deepseek-r1), with three runs per patient record. Outputs were compared with two independent human auditors against a clinician-adjudicated reference standard. Cloud-based models had higher observed F1 scores than the local models across all three domains. GPT-4o achieved an F1 score of 0.95 for preoperative symptoms; GPT-4o, o3-mini, and Gemini 2.0 Flash reached F1 scores of 0.65–0.68 for postoperative complications; and Sonnet 3.5 achieved an F1 score of 0.95 for disease course. These findings suggest that zero-shot LLM prompting could be a scalable approach for clinical information extraction while emphasizing persistent limitations in rare-event detection and privacy-sensitive deployment.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100327

Authors: Catalin Anghel Andreea Alexandra Anghel Antonio Stefan Balau Marian Viorel Craciun Adina Cocu Simona Moldovanu Cristian Sandu

Background: Multi-agent large language model (LLM) systems increasingly rely on structured interaction, debate, and final model-based evaluation, yet an unsupported causal attribution introduced by one participant may propagate through the interaction and influence subsequent decisions. This study introduces ClaimArena-LLM, a controlled framework for auditing adversarial propagation, evidentiary provenance, and final-stage evaluator vulnerability in multi-agent causal reasoning. Methods: ClaimArena-LLM uses 100 synthetic causal incident cases with predefined evidence-supported and false causal propositions. Each case was evaluated under matched adversarial and all-honest debate conditions using four functional roles: investigator, honest analyst, adversarial analyst, and auditor. The experimental program comprised 1460 validated runs, including the 200-run main experiment, claim-source lineage visibility and debate-round ablations, an 800-run balanced model–role rotation, and a 60-run proposition-identifier and presentation-order sensitivity control. Outcomes were scored deterministically from structured causal propositions, evidence references, claim lineage, reports, and auditor verdicts. Results: The adversarial analyst maintained the predefined false causal target with 97.6% mean fidelity. Adversarial exposure increased investigator shifts toward the false target by 16.0 percentage points (Holm-adjusted p = 0.035). At the auditor stage, false-target attribution increased by 11.0 percentage points (adjusted p = 0.043), verdict contamination by 20.0 percentage points (adjusted p = 0.002), and semantic provenance validity decreased by 20.0 percentage points (adjusted p = 0.002). Neither the manipulation nor the false target was explicitly detected. The ablations showed no robust adjusted differences, model–role rotation revealed substantial configuration dependence, and no benchmark-characteristic interaction remained significant after Holm correction. The proposition-identifier and presentation-order sensitivity control showed no statistically robust evidence that fixed proposition encoding explained the observed condition contrasts. Conclusions: Adversarial causal influence in multi-agent LLM systems is selective rather than uniformly propagative and depends strongly on functional role and model–role configuration. Within the tested architecture, a final LLM auditor cannot be assumed to provide an independent safeguard against upstream contamination. Reliable evaluation therefore requires separate assessment of causal correctness, propagation, semantic provenance, and final-stage adversarial awareness rather than reliance on final answer accuracy alone.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100326

Authors: Borislava Toleva

Model validation is essential for evaluating how reliably a machine learning model performs on unseen data, especially in imbalanced classification where performance estimates can be affected by an underrepresentation of the minority class. Cross-validation methods are widely used for this purpose. However, they do not always provide direct information about the quality of individual validation samples or the stability of model performance when the training data changes. This study proposes a reliability-controlled stratified bootstrap procedure for model validation in imbalanced classification. The method applies class-aware bootstrap sampling and constructs the test sample from observations not used for training. In addition, each bootstrap run is evaluated according to predefined reliability criteria related to minority-class representation, preservation of class proportions and train–test independence. Model performance is assessed using standard classification metrics together with minority-class measures, variability across bootstrap repetitions and the train–test performance gap. The proposed stratified bootstrap is evaluated across six imbalanced datasets and six classification models and is compared with holdout train/test splits, k-fold cross-validation, stratified k-fold cross-validation and stratified shuffle splitting. The results show that the proposed bootstrap produces performance metrics consistent with the benchmark resampling procedures while providing additional information about validation reliability and model stability. In addition, the parallel implementation of the proposed bootstrap reduces computational time in many of the experiments, although the computational advantage depends on the dataset and classifier. Therefore, the proposed bootstrap can complement established resampling procedures when reliability control, minority-class evaluation and stability under repeated changes in the training data are important.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100325

Authors: Mena Hanna Albalawi Al-Azani Kurdi Albadi

The rapid adoption of multimodal learning systems across high-stakes applications has intensified concerns about fairness and social bias, yet existing research remains scattered across modalities, tasks, protected attributes, mitigation strategies, and evaluation methodologies. This survey provides an overview of fairness and social bias in multimodal artificial intelligence. By reviewing 44 papers and 16 datasets, we synthesize the current landscape of fairness research in multimodal systems and identify emerging challenges and research opportunities. Our primary contribution is twofold. First, we introduce a multi-dimensional taxonomy that organizes the literature across nine operational dimensions, providing a unified perspective on multimodal fairness research. Second, we propose an end-to-end operational lifecycle framework that illustrates how biases emerge within shared latent spaces and identifies targeted pre-processing, training-time, and post-training intervention strategies. Overall, this survey discusses key performance-fairness trade-offs and provides a structured roadmap for evaluating and mitigating bias, ultimately supporting the development of more equitable multimodal AI systems.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100324

Authors: Jarek Tkocz Marek Sikora

In multi-stage manufacturing, conformity is confirmed only at the final test, after the cost has been incurred. On an automotive assembly line (16,500 workpieces), a regression model per checkpoint predicts the final measurement; a threshold turns it into a Go/NoGo decision. On the holdout, the models ranked the defects (area under the curve, AUC, up to 0.81), yet at the inherited final-station limit they raised no alarm and missed all 15. The threshold was therefore calibrated per checkpoint on an earlier period under a cost-weighted objective with an explicit defect rate, and evaluated on a later holdout. Carried over as a value, it failed: 31% of production flagged during calibration became 89% on the holdout. Carried over as an alarm budget read off recent predictions and set from inspection capacity, it recovered 14 of 15 defects at a 36% false-alarm rate, below the cost of all nine baselines on this holdout; the cost objective at the defect rate the holdout later showed, a retrospective sensitivity scenario, gave the same count. Under 20 other seeds, the advantage over the closest rules did not hold, and a flag on the ramped-up variant, available beforehand, would have been cheaper. The evidence is a retrospective, single-line case study with 15 holdout defects; the rule is a candidate for shadow-mode evaluation, not a validated deployment policy.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100323

Authors: Jessica Velasco Melvin Cabatuan Argel Bandala Edwin Sybingco Cesar Llorente Rennan Baldovino Laurence Gan Lim James Manuel Medalla Stephen Wong Justin Ryan Tan

Automated binary semantic segmentation of focal liver lesions (FLLs) in multi-phasic computed tomography (CT) scans is important for knowledge extraction of hepatic malignancy, staging disease, and planning treatment. However, standard models struggle due to extreme class imbalances, complex boundary transitions, and tissue heterogeneity across phases. We present a systematic ablation study evaluating CNN Only, Transformer Only, and parallel dual-encoder networks that combine a local convolutional neural network (CNN) branch (VGG-19 or ConvNeXt-Tiny) with a global hierarchical vision transformer branch (Swin or CSwin) against standard and state-of-the-art baselines for semantic segmentation in medical images. These backbones were evaluated across feature fusions as: early, intermediate, and late strategies on a 308-patient subset of the MCT-LTDiag dataset, equally partitioned across five FLL subtypes (BCLM, CRLM, HCC, HH, and ICC). The study evaluates binary lesion segmentation (Foreground vs. Background Liver) across 5 distinct clinical subtype cohorts (BCLM, CRLM, HCC, HH, and ICC). Evaluation was conducted via patient-grouped, stratified three-fold cross-validation. Pairwise architectural comparisons were performed using a patient-level paired design, with statistical significance evaluated via bootstrap 95% confidence intervals (10,000 resamples), two-sided paired permutation tests (10,000 sign-flips), and Cohen’s du effect sizes adjusted using the Holm-Bonferroni method. The proposed dual-encoder configuration utilizing ConvNeXt-Tiny + Swin Transformer with intermediate feature fusion achieved an optimal trade-off in segmentation performance, yielding a top Mean Dice score of 0.875 ± 0.076 and a Mean IoU of 0.861 ± 0.082. Compared to the best CNN-only baseline (ConvNeXt-Tiny), the proposed model showed statistically significant improvements across all parameters: +0.032 Mean Dice, −15.00 mm HD95, and +0.137 Normalized Surface Dice (NSD). From a resource-complexity perspective, the optimal ConvNeXt-Tiny + Swin Transformer with Intermediate Fusion model operated at <8.1% of the GFLOPs (33.34 vs. 412.30 GFLOPs) and achieved ~6.7 times faster with inference speedup (24.66 ms vs. 165.00 ms latency) than computationally intensive state-of-the-art 2D nnU-Net baselines.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100322

Authors: Aleksandr Volkov Taras Pustovoy Dmitriy Istomin Tatiana Otbetkina

Retrieval-augmented generation and structured agent memory are commonly treated as competing ways to ground large language models. Our results show that their relative value changes with the task and that neither retrieved similarity nor graph association is sufficient evidence for an answer. We present a task-adaptive two-layer methodology combining a compact, reviewed knowledge layer for orientation and cross-source synthesis with a versioned evidence layer for exact lookup and claim support. A router selects the representation, while an independent admission stage checks named entities, time, source role, contradiction, and provenance. A controlled course-grounding study reveals a task-conditional crossover within Course 2: the distilled Body of Knowledge (BoK) trails plain RAG by 1.07 points on lookup and leads by 0.62 points on synthesis. Across a 72-question three-course lookup panel, RAG leads by 2.17 points; Course 1 synthesis is near parity. Complementary, position-debiased LLM-judged mastery comparisons favor the BoK on both courses. The standing representation is reduced 22–35-fold. An observational production case over a 289,642-chunk evidence base illustrates why candidate discovery, packet selection, and evidence admission require separate gates. The contribution is a tested method for assigning retrieval, structure, and evidence admission distinct responsibilities within one grounded agent architecture.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100321

Authors: Horacio Thompson Rocío A. Conforti Sebastián A. Andujar Marilina Casais Carlos M. Telleria Marcelo L. Errecalde

The reciprocal relationship between large language models (LLMs) and knowledge graphs (KGs) opens new opportunities for knowledge discovery from scientific literature; however, most approaches address only one direction of this interaction. In complex biomedical domains, ensuring reliability, interpretability, and traceability remains challenging, making expert supervision essential. In this work, we propose a framework that integrates LLMs, KGs, and a Human-in-the-Loop scheme for constructing, evaluating, and reasoning over Evidence-based Knowledge Graphs (E-KGs), explicitly linking each extracted triplet to supporting textual evidence, and demonstrate its application to study the phenomenon of resistance in ovarian cancer. The E-KG was constructed using MedGemma-27B, with experts participating throughout the E-KG lifecycle, supported by our interactive visualization tools. We evaluated the E-KG through an expert-driven protocol assessing coherence and biological plausibility with inter-annotator agreement (Gwet’s AC1 coefficient), complemented by GraphRAG-based tasks, including link prediction (Hits@k, MRR) and a question-answering case study (faithfulness and answer relevance by LLM-as-a-judge and expert audit). The framework produced a coherent, plausible, and auditable E-KG, highlighting the potential of evidence-grounded knowledge to support LLM-based reasoning. To our knowledge, we introduce the first E-KG specialized in ovarian cancer resistance, providing an expert-validated resource that advances trustworthy and explainable AI for biomedical knowledge discovery.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100320

Authors: Manpinder Singh Panesar Aman Kataria Sita Rani Aditi Kataria Vikram Puri

Compressed-TinyML pipelines apply distillation, quantization and pruning to fit lightweight anomaly detectors onto resource-constrained sensor nodes. Such frameworks are routinely called privacy-aware on the strength of on-device inference or noisy federated updates rather than any measured guarantee, yet that assumption is rarely tested with a membership-inference attack. Across six multivariate time-series datasets (sensor, facility and finance) and four compact reconstruction-autoencoder architectures (84 experimental cells), a detector accurate enough to deploy (detection AUC 0.878) is vulnerable to a reconstruction-error membership-inference attack (attack AUC 0.688; true-positive rate 0.074 at a 1% false-positive rate). A gradient-noise defense (gradient clipping with additive Gaussian noise) reduces measured leakage to near chance (attack AUC 0.508 as scored, 0.517 oriented away from chance, not exact chance; TPR cut 4.8-fold) at a utility cost of 0.054 in detection AUC (paired Cohen’s d=4.06 for the leakage drop, bootstrap 95% CI [0.145,0.210]), consistently across all datasets (one-sided paired Wilcoxon p=0.0156), while the utility change is nominal (p=0.031, d=1.36). Exported and executed as a real INT8 ONNX artifact, the model shrinks 3.2× but shows no evidence of improved privacy (mean leakage rises slightly; improvement test p=0.95, worsening test p=0.078, neither significant) and runs slower on average on a desktop CPU, so quantization is a storage optimization only. The primary attack is the loss-threshold attack; a shadow-model check on a subset reaches the same qualitative conclusion. The defense is not formal differential privacy, but the finding is clear: on-device processing does not imply privacy; the privacy of a compressed sensor detector must be measured, not assumed.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100319

Authors: Roberto S. Porto Solano Antonio Berlanga de Jesús José M. Molina Yair E. Rivera Julio

High-stakes and regulated decision-support systems require predictive models that are simultaneously accurate and interpretable, yet accuracy is frequently obtained through opaque models whose behaviour is difficult to audit. This paper develops an explainable-by-design classification framework in which interpretability is treated as a constructive objective rather than a post hoc add-on. We aggregate thirteen complementary relevance metrics into a single, stable ranking and then cast the selection of a minimal relevant feature subset as a multi-objective optimization problem that trades predictive quality against structural complexity. The resulting accuracy–complexity Pareto front is approximated with the NSGA-II algorithm, and an equivalence-aware rule selects, among statistically indistinguishable models, the most parsimonious and hence most transparent one. We further (i) characterize the feature-subset ε-Rashomon set of each problem, (ii) quantify the stability of the aggregated ranking under resampling and contrast it with the individual metrics, and (iii) benchmark the induced subsets against established selectors (mRMR, Boruta, RFE, LASSO) and against opaque full-feature baselines, including gradient boosting. The framework is validated with a leakage-free nested cross-validation—in which the ranking, the search and the selection are recomputed inside each training fold—across six heterogeneous domains (healthcare, industrial production, climate, socio-economics, and education). Empirically, reducing the feature space to about three features preserves predictive accuracy: under paired tests corrected for fold dependence (Nadeau–Bengio) and for multiple comparisons (Holm), the minimal model is statistically indistinguishable from the full-feature model on all six datasets and is never significantly worse: reducing to a handful of features is essentially free. The main payoff is therefore not an accuracy gain but the characterization it enables—a simple ranked prefix performs comparably to the equivalence-aware selection (a direct signature of the Rashomon effect), and every problem admits a large set of near-equivalent minimal subsets sharing a compact stable core, which we map explicitly. The aggregated ranking is also more stable than the average individual metric in four of six datasets. These results support an explainability-by-design paradigm in which the accuracy cost of transparency—incurred only on the hardest multiclass tasks—is small, explicit, and justified when auditability and human oversight are required.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100318

Authors: Petra Radočaj Dorijan Radočaj

Zebra finches (Taeniopygia guttata) possess a diverse vocal repertoire that plays an important role in their social communication, yet automated classification of their vocalizations remains largely unexplored. This study presents a biologically informed multi-task learning framework that jointly classifies zebra finch vocalizations at two hierarchical levels: the call-type level, comprising eleven individual call types, and the semantic level, comprising six broader categories derived from an established perceptual taxonomy. Frozen embeddings from the BirdAVES pre-trained audio transformer are combined with a shared encoder trained under a composite objective integrating call-type cross-entropy, cross-entropy over semantic targets, and Supervised Contrastive Loss. Four experimental configurations are evaluated, comprising two single-task baselines, one for call-type and one for semantic classification, and two multi-task variants differing in their semantic loss weighting strategy. Across the two multi-task configurations, the best call-type performance is achieved by the dynamic weighting variant, reaching 87.11% accuracy and a macro F1-score of 82.76%, while the best semantic performance is achieved by the scheduled weighting variant, reaching 90.55% accuracy and a macro F1-score of 90.74%. Both multi-task configurations outperform their respective single-task baselines. Embedding space analysis reveals an approximately three-fold increase in silhouette score under multi-task learning, suggesting that biologically informed supervision may improve the geometric structure of learned acoustic representations in a manner consistent with the functional organization of the zebra finch vocal repertoire.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100317

Authors: Warodom Phungjununt Kanabadee Srisomboon Wilaiporn Lee Akara Prayote Luepol Pipanmekaporn

Biomedical question answering requires multi-hop reasoning over implicit relationships—e.g., identifying gene-associated diseases to retrieve target drugs. However, closed-book frontier language models achieve only ~14.57% precision on two-hop queries under single-answer prompting and answers drawn from parametric knowledge lack verifiable evidence, limiting their clinical utility. In this paper, we propose Bridge-Aware Reinforced Compositional Exploration and Decomposition (BRACED). A planner trained with Group-Relative Policy Optimization emits sub-queries, and a deterministic type-aware two-hop traversal over PrimeKG converts them into an answer set, so every returned entity is traceable to a graph path. On the BioHopR benchmark, BRACED reaches micro-F1 0.4241 and macro-F1 0.6381, improving over supervised fine-tuning of the same planner by 30.2%. With retrieval and the cap held fixed, five backbones of 1.7B–8B finish within 0.0034 of one another and BRACED reaches its result with one language-model call per question, from a 7B open-weight model that runs on the premises. Two typed hops already reach 95.1% of gold answers on this PrimeKG-derived benchmark: performance is bounded by candidate selection under the cap, not by decomposition quality.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100316

Authors: Pitchayanida Khumwichai Dussadee Praserttitipong Sayan Unankard Kornprom Pikulkaew

Occupation-aligned course recommendation requires connecting occupational requirements with relevant educational content. CASPER-O addresses this task through occupation-specific retrieval, ESCO-based queries (CASPER Query), and expert-grounded encoder adaptation (CASPER-CL). The corpus contained 11,938 occupation–course associations and 4536 unique courses across ten IT occupations; the expert-annotated benchmark comprised 1200 pairs across eight occupations. Ordinal Krippendorff’s alpha was 0.685. The data partitions were globally course-disjoint. Two complementary protocols evaluated fully judged held-out rankings and retrieval of known held-out relevant courses from larger occupation-specific candidate pools without treating unjudged courses as nonrelevant. Relative to pretrained encoders using CASPER Query, adaptation increased Held-out Relevant Recall@50 from 0.2417 to 0.3208 for MPNet and from 0.3417 to 0.4292 for BGE, while nDCG@10 changed by +0.0037 and −0.0153, respectively. With query–course scores fixed, candidate restriction reduced candidate pairs by 75.53% while retaining all 46 known held-out relevant pairs. Sensitivity analysis found nof target mapping uniformly superior across backbones and metrics. Exact paired sign-flip tests did not attain significance after Holm correction; non-significance does not establish equivalence. Overall, the results distinguish candidate-space effects from model- and metric-dependent effects of query construction and expert-grounded adaptation.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100315

Authors: Andrei Dugăeșescu Adina Magda Florea

Artificial intelligence (AI) models, particularly those based on deep learning, have achieved strong performance across a wide range of real-world applications. However, as models increase in scale and complexity, their behavior and internal decision-making processes become increasingly difficult to understand. Explainable artificial intelligence (XAI) has emerged to address this challenge by developing methods for making model predictions and internal mechanisms more transparent and intelligible. Despite substantial progress, the XAI landscape remains difficult to navigate due to methodological diversity, inconsistent terminology, and increasingly distinct research directions. In particular, classical XAI techniques, which primarily explain model predictions and behavior, and mechanistic interpretability (MI), which seeks to understand internal representations and computations, are often studied separately. This paper provides an integrative survey of these approaches, bringing them together within a shared taxonomy that spans explanatory scope, stage, and methodological foundations. The review examines their underlying principles, assumptions, strengths, and limitations while also surveying evaluation criteria for assessing explanation quality and highlighting connections and distinctions across the broader interpretability landscape. By synthesizing these research directions within a common framework, the study aims to provide a cohesive reference for understanding, organizing, and comparing methods for explaining machine learning models.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100314

Authors: Gabriele Rescio Andrea Manni Andrea Caroppo Alessandro Leone

Gait Anomaly Detection is crucial for early diagnosis of neurodegenerative conditions and for monitoring rehabilitation progress. Surface electromyography provides a direct view into muscle synergy and motor intention, but the field suffers from a chronic scarcity of labeled pathological data. Traditional supervised methods, although accurate, rely on large, balanced datasets and often fail to generalize to unseen anomalies, limiting applicability in real-world clinical scenarios. An unsupervised framework based on a deep Autoencoder architecture is proposed. Instead of classifying anomalies, the model learns a robust representation of normal gait patterns through One-Class Classification, in a design that is inherently personalizable to each subject. Reconstruction errors are subsequently analyzed to identify deviations from normality, providing a sensitive measure of abnormal muscle coordination and motor patterns. The framework is validated on two distinct datasets: a controlled dataset of simulated anomalies, such as toe-walking, and a clinical dataset of real pathological gait patterns. Results demonstrate high detection accuracy: about 88% on real clinical knee pathologies and about 99% on the controlled setting of simulated toe-walking anomalies, the latter representing an easier, fully controlled scenario. These findings indicate that the framework effectively captures the muscle-activation synergies of normal gait and can detect deviations without requiring labeled pathological data. The unsupervised, reconstruction-based approach provides a practical, label-free solution for gait monitoring, with a normality model that can be adapted to the individual at deployment.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100313

Authors: Valentin Wiesner Frank Deinzer Alexander Rost

To improve physical processes, domain experts use experimental test setups to gather data. After sampling, the data is analysed using a regression model, for example. The objective of this work is to investigate the performance of different combinations of sampling and machine learning methods under varying process characteristics that are previously known to the domain expert. This results in practical recommendations for users with limited expertise in machine learning. A large-scale combinatorial test setup consisting of polynomial surrogates as pseudo processes with the characteristics of variable count, complexity, process noise and position of the optimum was created. The sampling methods (Sobol, random, Bayesian optimisation and response surface methodology with additional G-optimal test points) were compared against twenty-six ML methods. The results indicate that the choice of ML method has a greater influence on global modelling accuracy than the choice of sampling method. However, strong interactions between sampling and ML method demonstrate that they should not be selected independently. Overall, the findings suggest that process noise eliminates the predictive quality of standard regression models used in statistical evaluation. Space-filling sampling methods generally provide reliable performance, while regularised polynomial, kernel-based, and additive models offer strong predictive accuracy across multiple scenarios.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100312

Authors: Wenjie Xu You Wang Beilei Fan Yufeng Ren Chenxue Yang Xueqi Liu Yuting Wang Xian Li

Agricultural knowledge recommendation differs from standard item ranking because the desired output is often an actionable field plan rather than a document, a passage, or an isolated tip. A useful recommendation must match the crop, growth stage, problem, environment, and safety constraints while remaining traceable to technical evidence. We propose CEG, a claim-centric evidence graph framework that represents agricultural guidance as scenario-level recommendation cards. Offline, CEG converts heterogeneous agricultural materials into source-grounded claims, claim and entity relations, local recommendation items, and scenario plans. Online, CEG retrieves scenario plans first, falls back to item-level and raw evidence only when needed, and reranks candidates with scene fit, evidence strength, plan completeness, safety, and state-prior signals. The final output is a structured card containing diagnosis, ordered actions, execution conditions, risk controls, evidence references, evidence gaps, and next actions. On a frozen four-domain auto-silver benchmark with 100 examples per domain, CEG Full reaches 0.9327 macro Hit@1 and 0.9725 average MRR@5. The strongest ablation, CEG without plan synthesis, reaches 0.8830 and 0.9429, while PageIndex is the strongest standalone comparator at 0.8227 and 0.8631, respectively. We emphasize that these ranking metrics measure the ordering of designated relevant candidates only; they do not establish that generated plans are complete, safe, or satisfy all agronomic constraints.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100311

Authors: Dimitar Tanchev Alexander Marazov Gergana Balieva

African swine fever (ASF) remains a major challenge for veterinary surveillance because of its complex spatial spread across neighbouring regions. We present a Bayesian susceptible–infected–susceptible (SIS) model on a graph in which European NUTS2 regions are the nodes and transmission from a reported region j to a susceptible region i decays with geographic distance between their centroids, with strength pj→i∝dji−α1 and a small spontaneous-infection term. Transitions to a state with no reported ASF outbreaks are modelled using a logit-link function that combines regional case counts with continuing infection pressure from nearby infected regions. Bayesian inference is performed using stochastic variational inference (VI) and validated against Hamiltonian Monte Carlo using the No-U-Turn Sampler (NUTS). The model was evaluated on twelve hold-out months and compared with a persistence baseline. The proposed model achieved a lower overall log-loss than the persistence baseline (0.099 for VI and 0.095 for MCMC versus 0.127), with the largest gains on newly infected regions (information gain of +2.51 nats for VI and +2.65 nats for MCMC relative to persistence). VI and MCMC produced nearly identical predictive performance (average precision 0.895 versus 0.894), indicating that the mean-field variational approximation reliably captures the posterior predictive distribution for this model despite overdispersed marginal posteriors. For early warning of emerging outbreaks, flagging the ten highest-risk previously clear regions each month would have identified 36–38% of the 53 new outbreaks in the hold-out period, compared with none under the persistence baseline. The proposed framework provides an interpretable Bayesian model for regional ASF surveillance, demonstrating that distance-weighted graph-based epidemic models can improve early detection of emerging outbreaks.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100310

Authors: Jingxuan Li Weiping Ding Xingquan Zhu Guoqing Chao

With the advent of the big data era, data becomes increasingly important, especially for many machine learning tasks where data quality is vital. However, missing data is inevitable in many real-world applications, making its handling a challenging problem. Although some reviews exist, they either treat it as a statistical problem or focus only on traditional machine learning methods. This survey provides a novel taxonomy of missing data-handling methods. According to feature or label missing, we classify them into feature-missing handling methods and label-missing handling methods. For feature-missing handling methods, we further categorize them into four classes: traditional, deep learning, soft computing, and large language model missing data-handling methods. Each class is further split to introduce representative methods. For missing-label handling methods, we focus on label completion and distinguish active learning, which selects unavailable labels for acquisition from an oracle, from semi-supervised learning, which estimates unavailable labels from labeled and unlabeled data. For each category, we list their advantages, disadvantages, and applications. To facilitate engineers and researchers, some commonly used software is introduced. To promote further development, we point out challenges requiring deeper investigation.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100309

Authors: Vasileios Charilogis Ioannis G. Tsoulos Anna Maria Gianni

Continuous optimization now draws on a large and growing number of competing metaheuristic families, including differential-evolution lineages, covariance-matrix adaptation strategies, swarm-intelligence methods, and simulated annealing, yet no single method dominates across problem classes, a consequence of the No Free Lunch theorem. This motivates algorithm selection: choosing, for an unseen problem, the method most likely to perform best. Building on the OptimSolution framework, we assemble a headless, reproducible batch-execution pipeline coupled with Exploratory Landscape Analysis (ELA), enabling large-scale, statistically grounded benchmarking together with feature-based meta-learning. We evaluate a set of optimizers of comparable overall strength, drawn from distinct algorithmic families, across a broad benchmark suite spanning classical scalable functions and established competition test suites, under a consistent multi-run evaluation protocol. A Random Forest classifier trained on ELA features, validated under a Leave-One-Problem-Out cross-validation scheme, is shown to predict the best-performing optimizer for a previously unseen problem substantially more accurately than the naive Single Best Solver (SBS) baseline, closing a meaningful share of the gap toward the oracle Virtual Best Solver (VBS). Family-level analysis further shows that structured competition benchmarks and classical scalable functions tend to be won by different solvers, confirming that the underlying selection problem is genuine rather than trivial. These results indicate that landscape-aware selection can meaningfully outperform naive strategies when the candidate methods are drawn from heterogeneous algorithmic families of comparable strength.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100308

Authors: Nasser Khalili Alireza Saleh Sedghpour Mohammad Jahanbakht Zohre Hosseini

Credit-scoring systems must translate predictive risk into calibrated, uncertainty-aware, and governable decisions rather than ranking applicants alone. This study aims to develop and evaluate a governed credit-scoring architecture integrating predictive performance, probability quality, uncertainty, and governance within a unified decision-oriented framework. Existing studies typically evaluate these functions separately, leaving uncertainty and governance outside the core modeling pipeline. We present the Calibrated, Adaptive, Knowledge-aware Ensemble for Credit Scoring (CAKE-CS), a structure-aware architecture integrating fold-local dependency representation, heterogeneous experts, sparse routing, cross-fitted stacking, probability calibration, and cross-conformal decisions within a fold-disciplined pipeline. Across eleven public credit and default portfolios containing 70,518 observations, CAKE-CS achieves a macro-averaged Area Under Curve (AUC) of 0.869 versus 0.862 for the strongest gradient-boosting competitor, together with the lowest macro-averaged expected misclassification cost, Brier score, and log loss among nineteen competitors under the primary protocol. Bidirectional ablation identifies stacking as the clearest positive contributor to AUC, while routing and structure serve primarily architectural roles. Empirical cross-conformal coverage ranges from 0.882 to 0.920. The originality is architectural rather than algorithmic. CAKE-CS reconfigures credit scoring from a model-centric prediction problem into a governed decision-system problem by linking prediction, probability quality, uncertainty, action, monitoring, and audit evidence. Theoretically, this extends credit-scoring research by treating these dimensions as interconnected components of decision-system performance. Practically, the architecture provides a deployment-oriented template in which calibrated predictions and uncertainty support decision thresholds, human review, monitoring, and audit.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100307

Authors: Chika Obata Mizuki Dai Kenya Jin’no

Intrinsic-dimension (ID) estimates and active-unit diagnostics address different aspects of variational-autoencoder bottleneck selection. We evaluate search policies on four image datasets, five training seeds and five regularisation weights under explicit validation targets and common candidate costs. Holding reliability checks, fallback and reference charges fixed, ID is accepted only for 25 MNIST conditions. ID-derived and gated midpoint starts select identical widths, requesting 8.84 and 5.84 candidates on average; the other 75 conditions share an ascending fallback. In 10,000 coefficient–threshold replay evaluations, 155 local outputs miss the smallest qualifying width. A separate five-seed MNIST study adds Monte Carlo evidence-lower-bound capacity selection. Native pruning experiments distinguish full-model reconstruction, selected-axis masking, decoder adaptation and dimension transfer. On dSprites, zero-centred relevance-prior masking increases MSE by a mean seed-wise 128.6%; after equal additional decoder training, the unmasked model meets the quality target in all three seeds whereas the masked model remains 60.9% worse and fails it. Active-unit prediction establishes no additional benefit on its near-ceiling task. These bounded results separate representation geometry, which intrinsic dimension measures, from reconstruction-sufficient width, which also depends on objective, training budget and quality target, and support controlled evaluation of selection rules and diagnostics without a general efficiency or pruning guarantee.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100306

Authors: Dalibor Radovanovic Ivan Milovanovic Ali Elsadai Aleksandar Sandro Cvetkovic Petar Kresoja Marko Sarac

Physics-informed anomaly detectors for industrial control systems are usually validated on a single synthetic or testbed dataset, which can yield optimistic estimates that do not transfer to deployment-like conditions. This validation-to-deployment discrepancy, the reality gap, is normally treated as one phenomenon and attributed to generic domain shift. We argue that it comprises at least three mechanistically distinct failure modes and propose a taxonomy that separates them by what a practitioner can observe and act on: Type 1, cross-regime evaluation; Type 2, operating-point shift; and Type 3, temporal distribution shift. A single physics-informed framework combining a Kalman-filter validation engine, BiLSTM, β-VAE, and graph attention network with residual sharing is evaluated across HAI 23.05, SWaT, and WADI over five random seeds. The three cases behave differently. Type 1 shows a substantial cross-regime decrease: AI-only falls from 0.774 in the prior synthetic evaluation to 0.493 on HAI, and Full EADE from 0.798 to 0.496. Under the retrospective fixed-FPR comparison on HAI, GNN+BiLSTM has the highest mean F1 at 0.498, although it is not separable from Full_EADE across five seeds. Type 2 collapses to FPR = 0.850 with F1 ≈ 0.237 across configurations; a strongly shifted sensor gives the clearest localized signature, but the failure is driven by aggregate distributional change and is recoverable by global threshold recalibration. Type 3 shows moderate FPR (0.244) and the largest sensitivity to residual occlusion (ΔF1_occl = +0.086), although a trained residual-free ablation shows no corresponding loss on any dataset. The resulting taxonomy links distinct failure signatures to different deployment responses.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100305

Authors: Luis Vera Olga Craveiro Ricardo Malheiro Manuel Dias Ricardo Correia Bezerra

Large language models (LLMs) are increasingly proposed for clinical summarization, yet evaluations relying on textual similarity overlook clinically consequential failure modes such as fabrication, omission, contradiction, and negation inversion. We treat medical summarization as controlled clinical compression, operationalizing a multidimensional framework of seven quality dimensions, an eight-category error taxonomy, and fourteen replicability components. The prototype runs on N=511 de-identified Portuguese-language Picture Archiving and Communication System (PACS)-derived multi-modality imaging reports, generating English impressions via local gemma4:latest under three prompt variants (R03) and temperature control. Factuality was scored claim-by-claim by an LLM judge with a two-pass protocol yielding 100% label coverage on 6190 claims, cross-checked by three external judges and two non-radiologist physicians. Factuality reached S=99.38% (P1, 95% confidence interval (CI) 98.94–99.64), 99.54% (P2, 99.12–99.76), and 92.77% (P3, 91.60–93.79); critical errors were 0.50% (95% CI 0.35–0.71); critical omissions (source-to-summary) are outside the claim-level scheme. Pairwise Cohen’s κ among external judges ranged 0.677–0.908. A controlled Portuguese (PT) → PT re-run (n=50) yielded significantly lower factuality than PT → EN (Δ=−7 to −15 pp per variant, p<0.0001 paired), restricting claims to PT → EN. Human annotation elicited a factuality–completeness criterion gap; BERTScore F1 correlated weakly with factuality (r=0.047); temperature had no significant effect.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100304

Authors: Fatemeh Barati Mohammad Soltanian Keivan Borna Luca Longo

Growing deployment of Convolutional Neural Networks in safety-critical domains has intensified the need for transparent, discriminative explanations, yet gradient-based methods such as Grad-CAM compute explanations for a single class score, producing saliency maps that activate regions shared with competing alternatives. This limitation is particularly consequential in multi-class and fine-grained recognition settings, where non-discriminative explanations reduce the practical utility of the method and impede reliable model auditing. This research study contributes to the body of knowledge by proposing Contrastive Gradient-Flow Interpretation (CGFI), an operator on gradient-weighted attribution maps that applies the contrastive principle that discrimination is sharpened by modelling what a class is not, to generate class-discriminative explanatory maps. CGFI explicitly models the top-K competing classes, computes their aggregated gradient attribution, and subtracts their influence from the target class attribution at the feature-map level, emphasising regions uniquely characteristic of the predicted class. It was evaluated across four backbones (Grad-CAM, Grad-CAM++, XGrad-CAM, Score-CAM), with and without contrast, on ResNet-50, VGG-16 and VGG-19. CGFI reduced the target-rival rank correlation across all 1162 image-backbone combinations evaluated, and improved the target probability share in 11 of 12, while preserving attribution fidelity. These findings suggest that contrastive attribution is a principled mechanism for improving explanation specificity.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100303

Authors: Martin Pavlíček Lubomír Štěpánek Martin Kotyrba Eva Volná

We examine how developer origin, citizenship-perspective instructions, and English versus Mandarin presentation shape sentiment scores from large language models (LLMs) for nationality-matched actions. Across two experiments using the same eleven-model panel, we collected 4,065,600 response records: 2,032,800 for the main corpus and 2,032,800 for the earlier corpus. Each corpus contains 26,400 bilingual stimulus rows, comprising 13,200 nationality-matched pairs across 22 event categories, two intended polarities, and three temporal frames, evaluated under seven conditions. Inference about developer origin uses ten models from the United States (US) and China and a nine-developer sensitivity analysis. In the main experiment, the English baseline developer-origin gap is +0.698 score points (exact model-label permutation p=0.175; developer-level p=0.238). Citizenship-perspective instructions produce language-averaged attribution changes of +5.20 points for the US persona and −5.76 for the China persona. All ten models shift in the requested direction in both English persona conditions; nine do so in each Mandarin condition. Mandarin changes absolute scoring, especially for positive-design items, while its average factorial effect on the paired attribution contrast is small and uncertain. The temporal origin interaction is also uncertain (−0.142 points; p=0.587), despite polarity-dependent temporal patterns. Non-numeric returns account for 0.064% of main-experiment responses and occur only among hosted models. The experiments converge on substantial average persona steering and modest, uncertain baseline origin differences, with model-specific exceptions. Corpus content, persona wording, and collection period vary together, and reuse of the panel adds no independent origin units. The findings identify explicit perspective instructions as a consequential part of the scoring protocol within this fixed panel. The audit concerns nationality-matched generic scenarios and does not establish equivalence between origin groups or the absence of broader geopolitical bias.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100301

Authors: Doyinsola Olatinwo Mae Seto Bruce Martin Mark Thomas

Passive acoustic monitoring (PAM) has emerged as a non-intrusive method for detecting and classifying marine mammal vocalizations in near real time. Yet reliable classification from PAM remains challenging, particularly for whistles—transient, narrowband calls whose rapid temporal variations are difficult to capture from conventional spectral analyses alone. Consequently, existing approaches often rely on species-specific classifiers, which require extensive data collection, annotation, and computational resources, limiting their application across diverse taxa. To overcome these limitations, this study introduces a unified multi-feature framework that integrates (i) temporal, spectral, and cepstral (TSC) features and (ii) spectrogram representations for whistle classification across multiple marine mammal taxa. In addition, to fully exploit these complementary representations, a novel dual-stream architecture comprising a bidirectional long short-term memory (BiLSTM) stream and a time-distributed convolutional neural network (TCNN) + BiLSTM stream is designed to learn from both feature types, enabling richer characterization of whistle dynamics. The framework was evaluated using five-repetition group-aware Monte Carlo cross-validation (MCCV), with recording groups kept mutually exclusive across the training, validation, and test partitions within each repetition. Across beluga, dolphins, narwhal, killer whale, and pilot whale whistles, as well as noise, the proposed model achieved a macro F1 score of 0.944 ± 0.004, outperforming the LSTM (0.918 ± 0.009), CNN (0.712 ± 0.037), and ANN (0.538 ± 0.134) baselines. Paired statistical comparisons also showed higher accuracy for the proposed model than the three baselines across the five MCCV repetitions, with the corresponding differences remaining significant after Holm–Bonferroni correction. These results demonstrate the effectiveness of integrating complementary acoustic representations with bidirectional sequence modeling for robust multi-species whistle classification across unseen recording groups. The proposed framework provides an effective approach for marine mammal whistle classification in PAM applications.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100302

Authors: Normtawan Phuttaamart Sunisa Kunarak

Reliable mobility management is critical for high-mobility 5G/6G applications such as smart cities, intelligent transportation systems (ITS), and connected vehicles, where reactive 3GPP handover mechanisms become unreliable as user equipment (UE) velocity increases. This paper proposes a hybrid Bidirectional LSTM (Bi-LSTM) and Deep Reinforcement Learning (DRL) framework for AI-native predictive handover management: a Bi-LSTM encoder extracts a knowledge embedding and forecasts future network observations, forming a predictive state that a Dueling Double DQN (Dueling DDQN) with prioritized experience replay uses to select handover actions. A closed-loop, dual-frequency refresh mechanism periodically updates both modules from accumulated network experience, without manual retuning. The framework is evaluated via Monte Carlo simulation across Smart City, ITS, and Connected Vehicle scenarios and Urban, Suburban, and Mixed eployments. The proposed Bi-LSTM predictor reduces RMSE by 51.9–68.6% relative to a persistence baseline and by 8.4–25.6% relative to a Transformer predictor, with the largest gains under high mobility. End-to-end evaluation shows the framework achieves a throughput of 67.0 Mbps, a latency of 48.1 ms, a handover failure rate of 3.47%, and an average utility of 0.678, outperforming all ablation variants; disabling closed-loop refresh causes the largest degradation (throughput loss: 7.0%, higher failure rate: 33.1%). Generalization experiments show bounded degradation under unseen conditions, with closed-loop adaptation recovering utility from approximately 0.588 to 0.678 within 100 refresh epochs after a domain shift. These results show that integrating predictive knowledge extraction, adaptive decision-making, and continual refresh provides a robust, closed-loop architecture for AI-native mobility management for proactive 5G/6G mobility management.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100300

Authors: Chengyi Zhang Zi Ye Ziyang Wang

Public surgical segmentation datasets contain information about the operative view that extends beyond their pixel labels. RoboSurg-VQA organises this information as a closed-set visual question answering (VQA) benchmark while retaining the source of each answer. It contains 5632 frames from 29 EndoVis 2017/2018 source sequences and 11 task units, with answers derived from source metadata, mask measurements, or machine-generated candidate labels. Seven visual attributes were examined in a blinded 250-frame audit by two study-team reviewers working independently, with disagreements resolved by a third reviewer. We trained a shared model with frozen BiomedCLIP encoders and a 13-answer vocabulary to examine the contributions of image and question inputs. Against mask-derived and candidate labels on held-out sequences, the model achieved a seven-task mean fixed-label Macro-F1 of 0.540. Answer frequency and task-restricted image-only prediction scored 0.391 and 0.400, respectively. With unseen paraphrases, the score was 0.528. On 40 audited frames from held-out sequences, mean Macro-F1 across bleeding, image quality, and glare was 0.535 against the human reference, compared with 0.450 for answer frequency, a gain of 0.085 driven by bleeding. The audit identified attribute-specific label errors, with the greatest reviewer disagreement for glare.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100298

Authors: Abha Tewari Dhananjay Kalbande

Reported gains from multimodal fusion, feature-attribution explanations, and model-confidence scores are only as trustworthy as the evaluation protocol and preprocessing pipeline behind them, yet these are rarely audited together. We develop a benchmarking-and-auditing methodology addressing four failure points jointly, subject-overlapping evaluation, class-imbalance correction under a subject-wise split, feature attribution, and calibration, instantiated on seven models under leave-one-subject-out (LOSO) cross-validation spanning three encoder-backbone pairings and early, late, and hybrid fusion, using VARISHTA-MM50, a synchronised IMU-and-video dataset of 50 elderly participants (aged 60–90). Random Forest under late fusion achieves the best LOSO accuracy (85.00%), outperforming every deep architecture tested. Subject-overlapping evaluation overstates accuracy by a mean of 8.32 percentage points; a GroupKFold control isolates subject overlap as the cause, and the effect replicates on an independent dataset. A leakage-safe oversampling protocol confirms the model ranking is not a preprocessing artefact. TreeSHAP attribution shows individual inertial features outweigh individual video features (1.32× per feature), though the wider video branch contributes more in aggregate, and a modality-removal ablation shows this dependence is specific to the tree ensemble. The most accurate model is the least well-calibrated by Expected Calibration Error. Subject-independent benchmarking, leakage auditing, attribution, and calibration form one connected evaluation discipline rather than four separate checks.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100299

Authors: Vinicius de Carvalho Rispoli Alan Müller Weverson Garcia Medeiros Marília Miranda Forte Gomes Thiago Alves Martins Rodolfo Deusdará

Structured clinical variables may complement fundus photographs, but the most appropriate point at which to integrate them into a retinal foundation model remains uncertain. This study compares an image-only baseline with three strategies for adding age, sex, diabetes status and hypertension status to a DINOv3-based classifier, namely a manually encoded visual metadata patch, learned metadata tokens and late fusion by feature concatenation. DINOv3 ViT-B/16 and ViT-L/16 backbones were adapted with Low-Rank Adaptation under a matched training protocol and evaluated on three patient-wise folds of the Brazilian Multilabel Ophthalmological Dataset. The diabetic retinopathy task comprised 9531 images with 1070 positive cases and the hypertensive retinopathy task comprised 8745 images with 284 positive cases. For diabetic retinopathy with ViT-L/16, all three strategies improved the F1-score over the image-only model. Late fusion reached an ROC-AUC of 0.9943, an AUC-PR of 0.9700 and an F1-score of 0.9326; the visual patch had a similar AUC-PR and learned tokens had the highest sensitivity. For hypertensive retinopathy, late fusion had the highest mean ROC-AUC of 0.8970, AUC-PR of 0.4441 and F1-score of 0.4338, but its paired advantage over the image-only model was limited to threshold-dependent metrics. Removing the target-related systemic variable reduced the late-fusion F1-score from 0.9326 to 0.9228 for diabetic retinopathy and from 0.4338 to 0.3776 for hypertensive retinopathy, showing that part of the multimodal gain came from disease-related clinical context. Models using only clinical metadata performed substantially worse than the image-based models. Metadata integration was beneficial in several matched comparisons, but early integration was not uniformly superior and late fusion yielded the most consistently favorable mean results.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100297

Authors: Paola Andrea Mejia-Zuluaga León Dozal Juan C. Valdiviezo-Navarro

From the Environmental and Remote Sensing community, the need for AI models that are accurate but also interpretable is increasing. However, usually Deep and Machine Learning techniques are uninterpretable. In this study, we propose to use Genetic Programming (GP) because of the readable nature of candidate solutions, like arithmetic expressions, for the detection of Phoradendron velutinum on five-band multispectral aerial images. Nevertheless, GP suffers from the bloat problem, which is the uncontrolled size growth of the candidate solutions, making them uninterpretable. With the aim of obtaining smaller, simpler, and interpretable solutions, we propose to test five GP simplification methods. The methods considered were Gws 25, Pruning 2%, and Gws 25 and Pruning 2% together. In addition, we proposed two novel approaches based on the Rademacher Complexity (RC) metric called Rademacher and Fit Rademacher. These strategies were evaluated over 30 independent runs with five-fold cross-validation on 250 images of 300 × 300 pixels. The folds group tiles by source frame, so the reported metrics are within-site estimates. The Fit Rademacher method achieved a 54% reduction in the average tree size (22 Fit Rademacher vs. 48 nodes Plain GP) and the highest Recall among all methods (0.566), trading precision and overall agreement for that gain: relative to Plain GP, its test κw falls from 0.448 to 0.402, its F1 from 0.499 to 0.399, and its precision from 0.498 to 0.307. Among the solutions found, GP-VRI (GP Velutinum Red Index) is a compact expression that was algebraically reduced from 25 to 14 nodes, that uses only the Red (650 nm) and Blue (450 nm) bands, and whose structure is biochemically plausible in light of the flavanone, carotenoid, and chlorophyll content reported for P. velutinum. A second solution, GP-VSDI (GP Velutinum Spectral Difference Index), was algebraically reduced from 28 to 24 nodes and uses the same two bands: the R−B difference appears twice in its denominator, while its numerator reduces to a constant factor where R≥2B, independently confirming that this band combination is the most informative for distinguishing P. velutinum from its background on this dataset. These results establish RC-guided GP simplification as a knowledge extraction mechanism in the remote sensing problem studied here, producing models compact enough to be inspected and assessed by domain experts, whose structure is physically consistent with the known spectral behavior of the target species. The methodological contribution is the pair of RC-guided operators; the detection framework and dataset are inherited from previous work.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100296

Authors: Bernardo Ibarra Infante Remberto Sandoval Aréchiga Víktor Iván Rodríguez Abdalá Salvador Ibarra Delgado José Ricardo Gómez Rodríguez Ana Gabriela Castañeda Miranda

Network-on-Chip (NoC) fabrics are the dominant interconnect for many-core processors, yet widely used routing methods—dimension-order routing and turn-model-based adaptive variants—assume mostly local, low-variance congestion. Modern workloads generate dense and bursty traffic that breaks these assumptions, motivating data-driven routing. Following a PRISMA-ScR protocol across five databases (2015–2026), this survey reviews machine-learning-based NoC routing across a 48-study evidence base, where path selection is posed as a Markov Decision Process and addressed with reinforcement learning and supervised/unsupervised regression rather than fixed analytic rules. We contribute a tripartite taxonomy organizing prior work by learning paradigm—mapped onto four empirical classes: tabular Q-learning, deep reinforcement learning, hybrid approaches, and supervised CNN-based prediction—transmission medium (planar/3D electrical and photonic NoCs), and multi-objective evaluation criteria (latency, throughput, energy, thermal budget, aging, fault tolerance). We ground comparisons in cycle-accurate simulation, FPGA/ASIC synthesis and its hardware overheads, feature-engineering and data-collection pipelines, and the accuracy metrics (RMSE, R2, error %) reported for learned predictors. We identify three recurring gaps: limited scalability beyond small meshes, missing deadlock-freedom guarantees for learned policies, and over-reliance on synthetic traffic, and outline future directions including graph neural networks, federated learning, spiking models, and runtime Pareto-front optimization for deployable, verifiable routing.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100295

Authors: Faranak Hatami Mousa Moradi

Accurate prediction of drug-pair-associated polypharmacy side effects remains an important challenge in computational pharmacovigilance. Standard binary cross-entropy (BCE) may provide insufficient emphasis on difficult positive examples. We evaluated ClinicalFocal, an asymmetric focal-loss function that assigns different focusing exponents and class weights to observed positive and generated negative drug-side-effect-drug triples, within a relation-aware graph convolutional network. On TWOSIDES, five-fold cross-validation showed that ClinicalFocal increased accuracy from 0.699 to 0.892 (+19.3 percentage points), F1 score from 0.700 to 0.894 (+19.4 percentage points), AUROC from 0.766 to 0.914, and AUCPR from 0.714 to 0.860. Sensitivity increased from 0.702 to 0.909, while overall classification error decreased from 30.1% to 10.8%, a 64.1% relative reduction. Brier score decreased from 0.183 to 0.147, a 19.6% relative reduction, indicating improved probability-level predictive accuracy. After independent 100-trial loss-specific optimization, ClinicalFocal remained superior across all five folds. These findings support loss-function design as an important determinant of graph-based polypharmacy side-effect prediction.

Machine Learning and Knowledge Extraction doi: 10.3390/make8100294

Authors: Jessy Wu Brett J. Borghetti Anthony L. Franz Darren E. Holland Abigail A. Bickley

Unattended ground sensors (UGSs) are passive systems used to detect and classify nearby activity such as the movement of military vehicles or personnel. However, UGS performance is limited in complex environments. This work explores the application of deep learning to classify vehicles as either ‘heavy’ or ‘light’ using seismic, acoustic, and magnetic sensor data collected at varying distances from a road. Two approaches were evaluated: one using features from Short-Time Fourier Transforms (STFTs), and another using convolutional neural networks trained on time series data. Models were also tested under data partitioning strategies involving either spatial or temporal variation. Data fusion was examined by comparing models that utilized all sensor modalities to those using a single modality. Layer-wise Relevance Propagation was applied to identify informative modalities and frequency components. Models utilizing frequency features outperformed time series-based models, with a 3% accuracy difference between the top performers. A 6% higher accuracy observed with spatial-only data partitioning compared with temporal partitioning suggests an overestimation of performance, highlighting the need to incorporate temporally distinct data. The STFT-based fusion model achieved the highest accuracy. Acoustic data, particularly in the frequency range below 100 Hz, was most relevant. Overall, the results demonstrate the effectiveness of transformation of the time series data to the frequency domain and the importance of careful data partitioning for reliable performance assessment. Future work should explore alternative, finer-tuned neural networks.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090293

Authors: Tymoteusz Miller

Agentic artificial intelligence requires safeguards that remain effective across trajectories rather than only at individual decisions. This study introduces MARIS-TRA, a hierarchical temporal runtime assurance extension of Controlled Agentic AI Systems. A compact, auditable fragment of Signal Temporal Logic (STL) provides quantitative robustness for speed, arena, pairwise separation, restricted-zone, and bounded-recovery requirements. The final architecture uses invariant hard-safety formulas to define feasibility, while recovery/liveness is monitored separately and may trigger escalation. Across a 5750-episode main campaign, MARIS-TRA achieved 100.00% realized hard-safety window satisfaction, 100.00% hard-safety episode satisfaction (450/450; Wilson 95% CI 99.15–100.00%), zero collision episodes, and 2.55 ms mean latency in the core comparison. Under the primary strict numerical semantics, an independent CBF-QP baseline achieved 88.89% hard-safety episode satisfaction; all 50 strict failures were very small arena-boundary overshoots in the boundary-stress scenario, with no collision or separation failures, and the post hoc tolerance sensitivity reached 100% at epsilon = 10−4 normalized simulator units. An additional 8640-episode targeted validation examined recovery hysteresis, model mismatch, AHO control flow, and safety–recovery conflicts. Under confirmatory high-density testing, safety-only shielding preserved hard safety in 810/810 episodes, whereas joint-hard enforcement produced 28/810 separation-safety failures (3.46%; Wilson 95% CI 2.40–4.95%) without collisions. Actuation-noise and delay experiments further show that the formal result is a conditional predicted-trace certification rather than a disturbance-robust guarantee on future receding-horizon execution. The empirical claims are, therefore, limited to the evaluated continuous-action multi-agent setting, while the architecture remains policy-separable and auditable.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090292

Authors: Ehtesham Hashmi Sule Yildirim Yayilgan

The rapid growth of social media has intensified the spread of hate speech targeting individuals based on race, gender, religion, and sexual orientation, creating significant technical, social, and ethical challenges. To address this, we propose a comprehensive English–Norwegian framework that unifies multi-class, multi-label, and multi-task learning for hate speech detection while embedding fairness, transparency, and accountability. The framework comprises six phases, including ethical system design, English–Norwegian data collection, bias analysis and mitigation, explainable AI, and stakeholder validation. A 14K-instance English–Norwegian dataset was constructed from public social media posts and enriched with identity-specific attributes (religion, race, gender, and sexual orientation) using human-in-the-loop and model-assisted annotation. The modeling approach applies multi-label binary fine-tuning to detect potentially hateful, offensive, aggressive, and neutral content, alongside multi-task learning for mutually exclusive demographic attributes. To promote ethical AI, identity-related prediction sensitivity is addressed through synonym-based augmentation, identity-conditioned label-contrast augmentation, and balanced loss re-weighting. Transparency is ensured using LIME to provide token-level explanations of model decisions. On the combined held-out evaluation partition, NorBERT-small achieved a macro F1-score of 0.92 and multi-label subset accuracy of 0.94 for toxicity classification. Language-stratified analysis further yielded macro F1-scores of 0.92 for English-only, 0.89 for Norwegian-only, and 0.87 for English–Norwegian code-switched text, providing a more detailed assessment of performance across the different language conditions, and high aggregate F1-scores for social-attribute extraction; however, these aggregate results should be interpreted cautiously because several social-attribute classes are sparsely represented. Compared with the baseline model, the mitigation stage reduces prediction sensitivity to many of the evaluated identity substitutions, although residual disparities remain for some identity pairs. The Machine Learning (ML) and Deep Learning (DL) models are included as conventional reference baselines. Because the evaluated model families employ different architectures and task-specific training procedures, the cross-family results are interpreted descriptively rather than as evidence of the intrinsic superiority of any particular model family. The reported findings are specific to the evaluated corpus; cross-platform, cross-domain, and external-dataset generalization were not evaluated and are therefore not claimed.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090291

Authors: Angel Casas-Ordaz Diego Oliva Marco Pérez-Cisneros Guillermo Sosa-Gómez Itzel Aranguren Arturo Valdivia-G

In machine learning, feature selection is a fundamental component that plays a crucial role in improving prediction accuracy and reducing the computational time of classification models. Feature selection involves eliminating irrelevant or redundant features that can negatively affect model learning. This article proposes a wrapper-based method that uses a binary evolutionary algorithm for feature selection. The Differential Evolution algorithm performs feature selection and is combined with Mahalanobis distance and Cholesky decomposition (MaCRO-DE). Using Cholesky decomposition to calculate Mahalanobis distance helps to determine the proximity of feature subsets, which is useful for deciding which features to select or discard. Subsequently, two classifiers (K-nearest neighbors and the support vector machine) are used as evaluators, along with a V-shaped transfer function, to identify the most suitable option for the proposed method. To evaluate the effectiveness of the proposed methodology, the CEC-2017 benchmark and twelve widely used datasets from the UCI repository are employed. Experimental tests are conducted using thirteen prominent metaheuristic algorithms. The results indicate that the proposal effectively achieves its objectives, outperforming the benchmark across the performance indicators used for comparison, including Accuracy, F1-Score, Recall, Precision, Fitness value, and number of selected features.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090290

Authors: Vafali Soltanmuradov Riccardo Berta Luca Lazzaroni David Martín Gómez Alessandro Pighetti Francesco Bellotti

Certified robustness provides guarantees that a model’s prediction is stable within a specified perturbation region, but certified-radius assessment for Transformer text classifiers is computationally expensive because sentence-level certificates require repeated radius searches over multiple perturbation positions. Existing approaches primarily improve the precision or runtime of individual verifier calls, whereas repeated position-level search is an additional source of computational cost. We address this bottleneck with verifier-aware perturbation-position scheduling: selected positions undergo full radius search to obtain a candidate sentence-level radius, skipped positions are validated at that radius, and only failed validations receive fallback search. Thus, scheduling reduces repeated search effort but preserves verifier coverage of all eligible perturbation tasks. Experiments include an expanded 100-example SST-2 evaluation and a 30-example Yelp evaluation, together with one-token and two-token embedding-space perturbations, early-exit verifier cascades, and scheduling ablations, to assess certificate preservation and verification efficiency across different settings. On the expanded SST-2 all-position evaluation, the scheduler reproduces the exhaustive sentence-level radius on all examples, while substantially reduces the number of full radius searches. Additional sensitivity experiments show similar scheduling behavior under the evaluated depth and longer-sequence settings. The presented results suggest that perturbation-position scheduling provides a verifier-compatible way to reduce repeated work in certified robustness assessment. Next research steps should expand the evaluation by addressing large pretrained language models, alternative verification backends, and broader discrete text perturbation settings.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090289

Authors: Simona Juvină Ana Neacșu Nicolae

Graph-based sleep-staging models use different node representations, making it difficult to separate the contribution of the frontend from that of the graph architecture. We investigate whether a common raw-waveform encoder can improve heterogeneous graph models without model- or dataset-specific architectural tuning. We replace the Native frontends of five graph backbones with the same compact one-dimensional convolutional neural network (CNN). The CNN learns node representations directly from 30 s polysomnographic waveforms, while each model retains its original graph-construction mechanism, temporal context, and prediction modules. Three training strategies are compared across five datasets using matched five-fold subject-wise cross-validation: two-stage training (CNN–2S), end-to-end training (CNN–E2E), and pretraining followed by fine-tuning (CNN–PTFT). At least one CNN strategy outperforms the corresponding Native frontend in 22 of the 25 backbone–dataset combinations. Mean macro-F1 gains are 2.99, 2.34, and 2.53 percentage points for CNN–2S, CNN–E2E, and CNN–PTFT, respectively. Gains are positive on average for all sleep stages and are largest for N1 and REM. CNN–2S gives the strongest overall results and the most favorable aggregate performance–cost balance, although the effect on training time varies across backbones. The frozen CNN–2S representation also remains effective when transferred across datasets. The findings support node-representation learning as an important component of graph-based sleep-staging design.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090288

Authors: Tomohiro Hayashida Shinya Sekizaki Natsuki Kato

Reinforcement-learning-based pedestrian agents can acquire adaptive behaviors without hand-crafted motion rules, but training from scratch in each environment is computationally expensive and often environment dependent. This paper investigates a two-stage transfer-learning framework for Deep Q-Network (DQN)-based pedestrian agents and analyzes how single-agent pretraining changes subsequent learning and behavior formation in multi-agent environments. A shared Q-network is first pretrained over four single-agent source tasks using Conservative Q-Learning with adaptive source-task allocation and is then transferred as the initial network of every agent in four multi-agent target tasks. Unlike fixed curriculum training, the source tasks are not traversed in a predetermined order; instead, experience collection is reallocated according to current collision and goal-progress performance. Compared with random initialization, the transferred network improves early-stage learning in intersection, bi-flow, and bi-door environments and significantly reduces target-task training time in these three tasks. In the bottleneck environment, the improvement is limited and the reduction in target-task training time is not statistically significant. Behavioral analysis further shows that pretraining changes early exploration: in the bi-door task, transferred agents reach goal-side and door regions earlier while spending less time near walls. Post-training trajectories also reveal collision avoidance and diverse route choices. These results clarify both the benefits and limitations of single-agent pretraining for efficient and interpretable behavior acquisition in multi-agent pedestrian simulation.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090287

Authors: Hector Alan de la Fuente-Anaya Miguel Morales-Sandoval Heidy Marisol Marin-Castro

Process mining is a data-driven technique that acts as a bridge between data science and process management. One of its main tasks is enabling the identification of process models from event logs. However, when event logs contain sensitive data or compliance with data protection regulations is required, privacy concerns arise. To address these issues, we propose a novel process discovery method that ensures data confidentiality under an honest-but-curious scenario. Our method is the first to leverage Ring Learning With Errors (RLWE)-based encryption for process discovery, enabling computations over quantum-resistant encrypted data. Unlike existing cryptographic approaches for Privacy-Preserving Process Discovery (PPPD), the proposed method uses fully homomorphic and RLWE-based encryption to execute the required computations over ciphertexts while preserving the quality of the discovered process models. This leads to a significant reduction in execution time compared to other PPPD approaches. We evaluated our method using real-life event logs commonly employed to evaluate process discovery algorithms. The results demonstrate improved performance and enhanced privacy compared to prior approaches. This contribution represents a substantial advancement in PPPD, laying the foundation for a secure and feasible cryptographic framework.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090286

Authors: Pei-Chen Kuo Yen-Ching Chang

Accurate localization of the upper end vertebra (UEV), lower end vertebra (LEV), and apex vertebra (AV) is essential for automated scoliosis assessment and Cobb angle measurement. This study proposes a geometry-aware, multi-task sequence learning framework for vertebral landmark localization and scoliosis type classification. The framework integrates YOLOv8-OBB-based feature extraction with a 1D Residual CNN-BiLSTM architecture and a displacement-guided regularization term, termed Geometric Loss (GeoLoss). Images and reference annotations were obtained from the open-source Spinal-AI2024 dataset, and a development set of 550 radiographs was used for model development and internal evaluation. Vertebral detections were transformed into 12-dimensional geometric sequences encoding orientation, displacement, shape, confidence, and curvature. Robustness was assessed using four train–validation–test partitioning strategies with repeated experiments and five-fold cross-validation, with the 75/15/10 partition showing the strongest mean internal performance. On a disjoint held-out set of 58 unseen images from the same dataset, the model achieved a Type Accuracy, EV Accuracy (±1), and EV Accuracy (±2) of 96.55%, 86.63%, and 94.19%, respectively, with an EV mean absolute error of 0.559 vertebral levels. These findings support geometry-aware sequence modeling as an effective approach for automated vertebral landmark localization and scoliosis assessment.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090285

Authors: Medet Ashimgaliyev Ainur Zhumadillayeva Miras Mussabek Nurbek Saparkhojayev Peiwu Qin Dusmat Zhamangarin

Predicting progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) requires models that represent both regional brain abnormalities and their evolution across repeated examinations. We developed a longitudinal graph neural network that integrates structural magnetic resonance imaging, FDG-PET, regional imaging biomarkers, and clinical covariates across irregular follow-up visits. The study included 614 participants with baseline MCI from the Alzheimer’s Disease Neuroimaging Initiative: 218 converters to AD within five years and 396 non-converters, with 2438 eligible longitudinal visits. Each visit was represented as an 82-node brain graph based on the Desikan–Killiany atlas. Node features combined a 128-dimensional multimodal convolutional embedding with four regional biomarkers. Graph-attention layers modelled spatial dependencies, a node-wise gated recurrent unit modelled longitudinal dependencies, and masked temporal self-attention aggregated variable-length visit sequences. Participants were divided at the subject level into development and held-out test sets, and hyperparameters were selected by five-fold cross-validation within the development set. On the held-out test set of 123 participants, the model reached an area under the receiver operating characteristic curve of 0.859 (95% CI 0.795–0.915), balanced accuracy of 0.805 (95% CI 0.736–0.862), sensitivity of 0.781, and specificity of 0.832. The AUC was numerically higher than that of the strongest baseline, a CNN–GRU sequence model, which reached 0.832 (95% CI 0.758–0.894); the paired AUC difference was 0.027 (95% CI 0.009–0.098), the unadjusted DeLong p-value was 0.026, and the Holm-adjusted p-value was 0.052, which was not significant at the conventional 0.05 threshold after correction for multiple comparisons. In ablation experiments, removing temporal modelling reduced the AUC to 0.818, and removing the spatial graph structure reduced it to 0.808, the two largest reductions observed. Integrated-gradient analysis placed the highest importance on hippocampal and entorhinal regions. Combining graph-based spatial modelling with recurrent longitudinal reasoning was associated with higher discrimination than sequence modelling alone, though this difference was not statistically significant after correction for multiple comparisons. Validation was restricted to a single research cohort (ADNI), and no independent external dataset was used; prospective external validation on an independent cohort is required before the model can be considered for clinical use Because FDG-PET was unavailable for 19.3% of visits, we report the headline result separately from a sensitivity analysis restricted to participants with complete FDG-PET at every visit (development set cross-validated AUC 0.891 vs. 0.874 for the full cohort with masked missing FDG-PET); multimodal performance should be read as cohort-dependent rather than as a single unconditional figure.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090284

Authors: Róbert Skapinyecz

The paper presents a novel large language model-based neighbourhood search method for production scheduling using discrete-event simulation (DES) as the modelling and evaluation engine. The combination of large language models (LLMs) and DES represents a relatively new area of research with significant potential. The current study aims to demonstrate the applicability of LLMs for the automated closed-loop optimisation of production systems, using the DES environment as a configurable modelling tool and an evaluation engine at the same time, from which the results and the related contextual information are extracted at the end of each simulation run to provide feedback and context for the LLM in its autonomous search for improved solutions. The architecture was implemented with the use of Mistral Small 3.2 (24B) as the LLM and Siemens Tecnomatix Plant Simulation as the DES environment. The results from the proposed approach were compared with those of a genetic algorithm (GA) native to the applied DES environment, and with those achieved using simulated annealing (SA). While the proposed method was generally outperformed by the GA and the SA in terms of final solution quality and computational cost, it demonstrated advantages in terms of simulation demand while also providing a highly adaptable optimisation framework with significant potential for improvement and wider generalisation in the future.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090283

Authors: Maksim Penkin Andrey Krylov

Medical image segmentation requires spatial transformations that remain effective across heterogeneous image statistics and boundary conditions. We introduce FunKAN, a KAN-inspired operator that acts directly on spatial feature maps: analytical Hermite basis maps are evaluated on an input-conditioned deformed grid, scored against learned channel-wise templates, mixed by a softmax over modes, and aggregated across channels. A Hilbert-space functional approximation result motivates the function-space viewpoint, while the exact finite FunKAN operator is treated separately and is not claimed to inherit a universal-approximation theorem. We embed FunKAN in a U-shaped architecture (U-FunKAN) and evaluate it under the validation-only selection protocol on BUSI, GlaS, and CVC-ClinicDB. The frozen model is tested once per seed after three-seed training. U-FunKAN improves mean test IoU over a same-protocol U-Net by 6.35 percentage points on BUSI and 1.85 points on CVC-ClinicDB, while GlaS performance is comparable, with U-FunKAN 0.50 points lower. U-FunKAN uses fewer parameters than the retrained U-Net (15.95 M vs. 31.03 M) but is substantially slower in measured GPU latency, so we do not claim universal computational efficiency. The results support dataset-dependent gains from functional spatial parameterization rather than uniform dominance.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090282

Authors: Arif Ali Rehman Enrique Nava Baro Pablo Otero

Training under extreme class imbalance (>1:100) remains an open problem in weakly supervised learning. The standard remedy—loss-level reweighting (focal loss, asymmetric loss, class-balanced loss)—is widely adopted, yet its behavior at extreme ratios in Multiple Instance Learning (MIL) is poorly understood. On digital breast tomosynthesis (attention-based pooling over frozen EfficientNet-B3 features), we study the optimization bounds under extreme bag-level MIL imbalance (1:251), intervening at two levels: the loss surface (via reweighting) and the gradient dynamics (via a novel q-calculus gradient modification using the Jackson q-derivative). All three reweighting strategies degrade classification relative to unweighted binary cross-entropy (BCE), monotonically, eliminating the loss surface as the bottleneck. Extended evaluation (n=20 seeds) shows q-calculus gradient smoothing matches vanilla BCE (p=0.632, Cohen’s d=0.003) despite provably reducing gradient variance, establishing an empirical ceiling on the optimization-achievable area under the precision-recall curve (AUPRC) of 0.0912; loss reweighting defines the floor at 0.055. Focal loss is additionally catastrophically miscalibrated (ECE > 0.44 vs. 0.036 for vanilla BCE), a collapse that persists under adaptive binning. In this regime, exceeding the ceiling points to the data-representation level, not the optimizer. We further identify ratio-invariant safety—non-degradation at any imbalance ratio, satisfied by vanilla BCE and q-calculus but violated by all reweighting methods—and give recommendations spanning moderate to extreme imbalance.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090281

Authors: Sotiris Zikas Katerina Gkirtzou Theodor Panagiotakopoulos Yiannis Kiouvrekis

Neurosymbolic AI (NeSy AI) seeks to integrate the strengths of symbolic reasoning with computational learning methods, addressing fundamental challenges of each paradigm in isolation. Existing surveys have primarily organized this growing body of research by architecture. The systematic evaluation of NeSy systems against the foundational questions about knowledge–learning interaction raised in the literature has received far less attention. This paper introduces a multi-axis analytical framework that combines the six-type taxonomy proposed by Kautz with four foundational dimensions derived from the open questions raised by van Harmelen: the mode of integration between symbolic and computational learning components, the use of symbolic priors for learning, the enforcement of symbolic constraints for safety and bias prevention, and the production of symbolic knowledge from learning. Complemented by a systematic reasoning categorization (deductive, inductive, abductive), this framework is applied to categorize and analyze 70 NeSy papers. The analysis reveals that while symbolic priors for learning are widely adopted in the surveyed corpus, symbolic constraints for safety and fairness remain significantly underexplored despite being among the most frequently cited motivations for NeSy research. Task-level abductive reasoning is virtually absent, appearing in only three of the 70 systems, all but one from 2026. Fully integrated architectures (Kautz Type 6) remain scarce and largely theoretical, and bidirectional knowledge–learning interaction is rare. Six concrete gaps are identified, providing specific directions for future research in neurosymbolic AI.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090280

Authors: Veronica Mendoza Ekaitz Zulueta Xabier Basogain Javier Peña-Ceballos Julen Carasa-Castaño

Humans acquire meaningful language by storing perceptual categories, category-word mappings, and conditional IF–THEN rules in rich, cross-domain, multimodal contexts. Crucially, structured cross-domain input that pairs visual context and language (text) appears to be fundamental to this process, enabling individuals to acquire, for example, the lexicon. The present study investigates whether Large Language Models (LLMs) can learn words when trained on structured cross-domain input rather than text-only exposure. This paper incorporates a controlled Fictitious-Animal Paradigm featuring 32 creatures and 32 pseudowords, divided into two phases: single-animal and dual-animal scenes. Single-animal contexts entailed storing categories and mappings, while dual-animal environments involved storing logical constraints and performing inference. The authors propose a computational model centred on four capacities for context-based word learning: storing perceptual categories, category-word mappings, and the Mutual Exclusivity rule (formalised as a conditional IF-THEN statement), and retrieving stored information to assign novel labels through inference. The evaluation demonstrates that structured cross-domain input enables LLMs to exhibit behaviour consistent with conditional IF–THEN rules, allowing them to infer and acquire novel words based on current visual contexts. These findings suggest that, with this input that integrates context and text, LLMs display adaptive, real-time human-like inferential reasoning in word learning.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090279

Authors: Danni Huang Litong Song Yue Chen Shuangjiang He Ruiqi Wang Hongyu Shen Weishen Chu

This study proposes an explainable artificial intelligence (XAI) engineering workflow for provider-level healthcare claim audit prioritization using SHAP-guided hybrid retrieval-augmented generation (RAG). The framework integrates provider-level claim aggregation, tree-based risk screening, SHAP explanation, exploratory group-level SHAP clustering, policy concept retrieval, and constrained large language model audit narrative generation. Experiments on a public Medicare provider fraud dataset use the dataset-provided PotentialFraud label as a weak audit prioritization label rather than a legal determination of fraud. The results show that reimbursement exposure, utilization duration, claim repetition, deductible patterns, and beneficiary case mix contribute to provider-level risk scores. Additional cross-validation, calibration, threshold, and scale-confounding analyses indicate that provider size and financial exposure are important confounders, while non-scale and contextual features also retain predictive information. Beyond prediction, the framework organizes local SHAP drivers into exploratory provider archetypes and maps explanation patterns to audit-relevant policy concepts. Compared with pure embedding retrieval, the SHAP-guided hybrid retriever increases policy concept diversity and explanation alignment, although these retrieval metrics do not replace independent expert audit validation. Because the public dataset does not include referral pathways, inter-facility transfers, provider–network relationships, or care-coordination records, the framework cannot determine whether utilization patterns are explained by clinically appropriate referrals, regional access constraints, or multi-level care pathways. Its current applicability is therefore limited to provider-level audit prioritization using the available claims and beneficiary variables. The proposed system is positioned as a reproducible engineering prototype for cautious, human-reviewed audit support rather than a comprehensive or automated fraud determination system.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090278

Authors: Duc Nghia Vu Thanh Huy Nguyen Duc Thi Vu Janos Demetrovics

Influence maximisation traditionally assumes that each activated neighbour contributes independently to the likelihood of a user adopting information, ignoring conjunctive synergies where a set of users must be active simultaneously to trigger another. Relational games provide a formal language for such coalitional dependencies, but the necessary influence hypergraph must be hand-crafted by domain experts, making them infeasible for large, dynamic social networks. We introduce the Neural Probabilistic Relational Game (N-PRG), a data-driven framework that automatically learns a probabilistic hypergraph of influence coalitions from cascade traces. A feed-forward neural network, trained via gradient descent to predict user activation, is interpreted using Deep SHAP to extract important set-level triggers. These are calibrated into a stochastic cascade model, the Probabilistic Relational Game (PRG), which generalises the Independent Cascade to set-based activation. We define the Minimal Reliable Seed Set problem, prove its NP-hardness even in the deterministic case, and establish that the expected influence function is monotone. We further demonstrate that, unlike the Independent Cascade model, the influence function under conjunctive (AND-type) hyperedges is in general not submodular, which precludes constant-factor approximation guarantees and motivates the use of greedy heuristics. Extensive experiments on synthetic data confirm that N-PRG successfully identifies coalitional interactions of size greater than one and achieves targeted out-of-sample coverage. Semi-synthetic experiments on Digg and Twitter network topologies demonstrate that N-PRG discovers seed sets up to 45% smaller than Independent Cascade baselines, while providing interpretable coalition pathways invisible to black-box methods. N-PRG thus unites the flexibility of gradient-descent learning with the structural rigour of relational games for influence analysis.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090277

Authors: Mohanad Alayedi Ahmad M. Jaradat

The Internet of Vehicles (IoV) is revolutionizing intelligent transportation systems by ubiquitous connectivity of vehicles, roadside infrastructure, pedestrians, edge/cloud platforms, and smart-city services. With the IoV evolving towards highly connected, autonomous and data-driven mobility ecosystems, it needs to meet challenging requirements for low latency, scalability, interoperability, security, privacy and trust. This paper presents a comprehensive cross-layer approach for intelligent, secure and privacy-preserving IoV systems. It is built upon an analytical framework and systematically studies the perception, communication, edge/cloud computing, blockchain-enabled trust and application layers of IoV technologies. In addition, the paper presents an in-depth review of the enabling techniques such as machine learning (ML), deep learning (DL), reinforcement learning (RL), federated learning (FL), blockchain, cybersecurity mechanisms, digital twins, edge computing, 6G integration, and resource allocation. Moreover, it discusses the interplay and trade-offs between intelligence, security, privacy, computation, latency, and scalability. The survey also covers other significant challenges like intrusion detection, decentralized authentication, privacy-preserving learning, blockchain overhead, semantic interoperability, post-quantum security, and standardized datasets. This study is intended to serve as a structured reference for the development of scalable, trustworthy, and intelligent IoV systems by highlighting state-of-the-art techniques, open research gaps, and future directions.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090276

Authors: Dalibor Radovanovic Dusan Markovic Petar Kresoja Aleksandar Sandro Cvetkovic Vesna Radojcic Marko Sarac Nikola Savanovic

Physics-informed anomaly detection for industrial control systems (ICSs) is usually validated on synthetic data. How much of that performance survives on real hardware-in-the-loop (HIL) data is an open question. We re-implement PhySec-Edge, a hybrid BiLSTM/β-VAE/GAT/Kalman pipeline, as a full architecture and evaluate it on HAI 23.05. The gap is consistent across five seeds. AI-only F1 falls from 0.774 (synthetic) to 0.525 (HAI); the full hybrid falls from 0.798 to 0.512. The components diverge. PVE residual sharing transfers but weakens on real data (ΔF1 = +0.010 ± 0.008 across five seeds against +0.017 synthetic; ΔMCC = +0.013, positive in every seed). The PVE NIS decision score fails outright (raw FPR = 0.482; a correction brings it to 0.013, but detection does not recover). Under the retrospective fixed-FPR comparison, the preferred configuration is GNN+BiLSTM (F1 = 0.530 ± 0.002), outperforming full-EADE fusion (0.512 ± 0.003). From this split we derive the information-vs-decision pathway distinction, a way to reason about component-level generalization in hybrid architectures. We then set out what it means for physics-informed ICS detector design and edge deployment.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090275

Authors: Brian Scavotto

When an autonomous language-model agent fails, a growing body of work hands the diagnosis to a large language model (LLM) that reads the trace and names the responsible step. We ask whether this attribution separates two causes, a localized corruption from one bad input and a distributed degradation across many steps. Using controlled fault injection across two agent tasks, four trace-completeness levels and ten judges, we find discrimination weak wherever the causes are demonstrably identifiable from the trace. Most judges apply one propensity, naming the step that produced the result, so their two per-class accuracies become readouts of one number whose margin stays near zero. Two frontier models produce results consistent with opposite ends of that axis, one asserting an inconsistency the trace does not contain, the other abstaining. On the realistic task four of six judges invert, naming a concrete step more often on the distributed cause than on the localized one. The limit is not incapacity. Where contradicting the trace takes only a comparison the same judge localizes the fault in fifty of fifty runs, against one of fifty where it would take arithmetic and zero of fifty where the trace is consistent, and a 2500-fold sweep of fault magnitude leaves the margin unchanged. A deterministic non-model rule separates the causes on the main-panel traces, where no judge under the standard prompt exceeds 0.16 at full detail; one prompt intervention lifts the margin to 0.34. We read this as a reliability caution for LLM-as-judge attribution across the models, prompts, and constructed tasks evaluated here.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090274

Authors: Kushagra Agrawal Deshmukh Nirmiti Akshay Palak Kaushik Shaveta Jain Ganga Sharma Sumendra Yogarayan

Large language models (LLMs) incur prohibitive computational costs when deployed as monolithic systems for multi-domain query processing. This paper proposes Accelerated LLM, a modular architecture that replaces a single general-purpose LLM with an ensemble of task-specialised small language models (SLMs) governed by a neural query router and a Mamdani fuzzy inference system. The router embeds each user query using a frozen sentence encoder and classifies it across four task domains—summarisation, translation, question answering, and text generation—routing confident queries directly to the corresponding SLM. Ambiguous queries are escalated to a three-input fuzzy logic system operating on Query Length, inter-Domain Overlap Score, and Classifier Confidence, enabling principled handling of imprecise inputs. A reinforcement-learning feedback loop, validated through a controlled pilot deployment, continuously refines the routing policy. The complete pipeline, including the sentence encoder, totals approximately 2.14 billion parameters—a 98.8% reduction relative to GPT-3.5 (175 B). The integration of fuzzy logic into the routing stage raises classification accuracy from 91.5% to 94.3% and reduces the hallucination rate to 9.8% (minor) and 6.4% (major). Evaluated on healthcare-augmented benchmarks against ChatGPT-3.5, Claude, Mistral 70B, and two contemporary compact models (GPT-4o-mini and Llama 3.1-8B-Instruct), Accelerated LLM achieves competitive or superior task-specific performance at a fraction of the parameter count. A small-scale pilot evaluation in the legal domain indicates that the routing and fuzzy logic components retain partial effectiveness beyond the primary healthcare setting, though full multi-domain validation remains future work.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090273

Authors: Egor Savchenko Anna Maslovskaya

Plant diseases, being a subject of interdisciplinary research, significantly reduce crop yield, quality, and economic returns, while the misidentification of pathogens often leads to ineffective treatments and may harm beneficial organisms and ecosystems. This work develops an approach for robust visual classification of plant diseases under limited and heterogeneous data based on multi-scale fractal texture descriptors integrated into a convolutional neural network. The proposed method employs wavelet transform modulus maxima to extract two complementary fractal characteristics, local fractal dimension and singularity spectrum width, from leaf images at several spatial scales. These descriptors form multi-channel fractal maps fed into a fractal attention module (FAM) inserted after the third stage of a ResNet-50 architecture. The FAM learns to emphasize spatial regions where fractal properties are most discriminative, while a parallel branch encodes global fractal statistics into an auxiliary vector combined with backbone features at the final classification layer. Experiments are conducted on a large heterogeneous collection of 11 public plant disease datasets under 5-shot, 50-shot, and full-scale training regimes. The fractal-augmented model raises classification accuracy from 57.06% to 67.73% on 5 shots and from 80.81% to 86.11% on 50 shots, red outperforming the plain ResNet-50 in these settings, converges within 1–2 epochs versus 25–40, and shows markedly better resilience to color distortions, random occlusions, and grayscale conversion in most cases. The generated attention maps provide spatially explicit explanations of the model’s decisions, increasing transparency for practical use. The proposed approach demonstrates that fractal analysis, embedded as a modulating signal inside a deep network, can serve as an efficient and interpretable inductive bias, which is particularly valuable under data scarcity and noisy agricultural imagery.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090272

Authors: Dana Tyulemissova Aigul Shaikhanova Oleksandr Kuznetsov Aigerim Sambetova Kainizhamal Iklassova Aisanim Sarsenbayeva

(1) Background: Information diffusion models developed for graph-based platforms such as Reddit and broadcast architectures such as Telegram identify temporal features—particularly the timing of peak spread—as dominant predictors of coverage. Whether these predictors generalise to platforms where content is distributed through algorithmic recommendation rather than social-graph contagion remains an open question. (2) Methods: We analyse the YouNiverse dataset, comprising 133,364 English-language YouTube channels observed weekly from January 2015 to September 2019 (18.9 million observations). We derive channel-level diffusion features—including time-to-peak, post-peak decay rate, diffusion volatility, and upload frequency—and train three machine learning models (Linear Regression, Random Forest, and LightGBM) on two tasks: predicting peak weekly view growth (regression) and identifying viral channels (classification). A single-feature naive baseline (subscriber count alone) establishes the marginal contribution of the broader feature set beyond subscriber count alone, and a temporal split experiment (training on channels peaking before 2018, testing on 2018–2019) assesses cross-temporal stability. Because subscriber count and subscriber rank are measured at the October 2019 crawl, this is a retrospective characterisation rather than a strict real-time forecasting design. (3) Results: LightGBM achieves R2=0.776 (5-fold CV: 0.778±0.003) compared with R2=0.548 for the naive baseline, a net gain of +0.228R2. Because subscriber rank and subscriber count are near-perfectly collinear, we interpret them jointly as a channel-size dimension (42.2% of total mean absolute SHAP attribution), rather than as independent effects. Time-to-peak ranks fourteenth (1.1%), in contrast to its dominant role on Reddit (r=0.995, rank #1). For virality classification, LightGBM achieves ROC-AUC =0.967. Under the temporal split, Random Forest (R2=0.703) outperforms LightGBM (R2=0.683), showing greater cross-temporal stability within this retrospective split. (4) Conclusions: Within the 2015–2019 data, the results are consistent with algorithmic recommendation weakening the relationship between temporal diffusion dynamics and coverage magnitude at the channel level. Time-to-peak is weakly informative in this setting, while generalisation to the current recommendation system requires validation on newer data.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090271

Authors: Ismail Lamaakal Chaymae Yahyati Yassine Maleh Khalid El Makkaoui Ibrahim Ouahbi

The integration of Tiny Machine Learning (TinyML) into human activity recognition (HAR) represents a paradigm shift in artificial intelligence, enabling real-time, efficient, and privacy-preserving analysis on resource-constrained edge devices. This paper presents a comprehensive review of TinyML for HAR, covering foundational concepts, methodologies, and applications. the review examines the state of the art of existing works and approaches that combine TinyML and HAR, providing a detailed review and comparison of models, algorithms, and frameworks. This comparison sheds light on the effectiveness and limitations of different methodologies. Key contributions include a systematic taxonomy of HAR systems leveraging TinyML, a detailed analysis of optimization techniques like pruning, quantization, and knowledge distillation, and insights into state-of-the-art frameworks and datasets. Challenges such as scalability, energy efficiency, and generalization to diverse environments are critically examined, alongside solutions like federated learning, multimodal data fusion, the integration of generative AI, neuromorphic hardware, 5G/6G and Internet Of Things (IoT) connectivity, which are highlighted as transformative enablers for advancing HAR applications. This review serves as a foundational resource for researchers and practitioners aiming to harness TinyML’s potential in activity recognition systems.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090270

Authors: Ioannis Kazlaris Konstantinos Diamantaras Efstathios Antoniou Charalampos Bratsas

This paper studies hallucinations in structured extraction using prompt-based Semantic Role Labeling (SRL) as a controlled case study. We implement a DSPy-based pipeline in which each prediction includes a generated rationale and a citation to a governing rule, allowing extraction errors to be aggregated by attributed rule. This diagnostic signal is used by a Rationale-Oriented (RO) optimizer to perform targeted, human-readable signature revisions. Experiments use a representative 1000-sentence pool drawn from the CoNLL-2012 test split, with a nested 200-sentence development subset used for optimization and the remaining 800 sentences reserved as an unseen subset. After signature selection, the role-specific signatures are fixed; on the disjoint unseen subset, they improve over their corresponding unoptimized baselines for all nine evaluated 5W + 1H-aligned roles, although the magnitude of improvement and the development-to-held-out gap vary substantially by role, with the sparse ARGM-PRP and ARGM-CAU roles generalizing markedly less well than their development scores suggest. Across the full 1000-sentence pool, the selected-role, gold-predicate aggregate reaches 67.04% strict micro-F1; this is a within-study summary and not a full CoNLL SRL score. In the held-out ARG0 analysis, extraction outcomes are non-uniformly associated with cited rules, while signature perturbations show that rule citations are better interpreted as observable diagnostic proxies than as faithful per-instance causal explanations. Compared against MIPROv2 on the same extraction model, RO reaches a comparably low commission-error operating point without the recall collapse MIPROv2 incurs. Two further ablations locate the source of the effect: a full nine-role cross-model transfer to Phi-4-reasoning shows RO signatures carrying across model families unevenly, with gains concentrated on the weak, sparse roles and negligible-to-slightly-negative where the target model is already competent, and with transferred precision capped by its coarser span-boundary behavior; and a matched same-model self-critique ablation, in which Qwen3-14B supplies its own diagnostic feedback across the same nine roles, tests how much of the refinement effect can be recovered without a stronger external critic. Seen together, these ablations separate cross-model signature portability from teacher-capacity effects and clarify which parts of RO depend on the diagnostic structure itself versus the critic used to revise the signatures. The primary condition retains a design-time dependence on a proprietary teacher, and signature optimization is centered on a single open-weight extraction model, Qwen3-14B.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090269

Authors: Nurgali Kadyrbek Madina Mansurova

Global attention combines long-range lexical access with semantic-relation retrieval, making whole-layer replacement lossy. We present Strata-HeadQuotient, which statically assigns each key–value (KV) head to full-history attention (GLOBAL), 1024-token attention (LOCAL), or local attention plus an event-scoped typed graph read (LOCAL_GRAPH). For a 24-layer, 16-head decoder, we audited 384 heads and all 6903 pairs among 118 candidates before constrained subset selection. The Q25 export localized 96/384 heads: 81 were LOCAL and 15 were LOCAL_GRAPH. On 470 document-disjoint 8192-token windows (3.85 million tokens), graph-disabled Q25 perplexity was 1.00146× dense (95% document-bootstrap interval: 1.00131–1.00161); the 4096–8191-token ratio was 1.00201. On an untouched 2000-case semantic test, typed execution was 1998/2000, versus 374/2000 for a capacity-matched untyped graph. Two fresh adaptation/selection runs reproduced both directions, although every matched-capacity marginal graph-effect interval included zero. Q25 retained 78.125% of token-KV bytes and 98.44% of dense full-sequence throughput; a tested Q30 export failed. Thus, interaction-aware localization preserved 8192-token language-model quality, while a separate typed path controlled event-role behavior. The experiment does not establish that graph computation caused perplexity preservation.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090268

Authors: Kishan Kartha Alex James

Laboratory experimentation is shaped by practical constraints, so computational frameworks built for offline settings do not transfer cleanly to an online laboratory routine. Two mismatches dominate. First, they assume on-demand access to high-fidelity facilities, whereas characterisation, fabrication, and testing are rarely economical one sample at a time. Second, experiments produce occasional catastrophic, non-Gaussian errors that a fixed noise model handles poorly. Here, we introduce a coupled scheduling-and-verification layer that wraps a multi-fidelity optimiser. A queue scheduler models per-session overheads and batches expensive measurements to amortise them; a Fidelity-Aware Verification Protocol repeats suspect observations and escalates unresolved ones to a higher fidelity. The two share the same cost-amortised queue, so a verification escalation is dispatched as just another queued sample, and anomaly handling reinforces batching rather than competing with it. On synthetic functions, the layer reduces overhead-incurring sessions by 28–43% and, under catastrophic outliers, cuts regret by 51–84%. On a real three-fidelity materials dataset, it improves cost-efficiency and robustness. In a live deployment on polydimethylsiloxane-mediated 2D material transfer, it reached Raman-confirmed monolayer graphene in 26 trials using only two Raman sessions, and those discovered parameters transferred another 2D material, monolayer molybdenum disulphide, on the first attempt.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090267

Authors: Sorin Liviu Jurj

Physics structure-informed neural networks (Ψ-NN) promise to carry known physical relations into compact models, but tiny machine learning (TinyML) deployment adds compression, finite precision, compilation, and hardware constraints that can change how those benefits appear in practice. We study this end to end for Burgers odd-in-x antisymmetry. A Ψ-NN preserves the encoded relation through clustering, 8-bit integer (INT8) quantization, compilation, and Cortex-M target-instruction-set emulation, achieving 6.0% relative L2 error and an antisymmetry residual of 0.006; none of seven evaluated compression baselines matches both under the common protocol. Deployment accounting reveals a complementary result. Centroid count is not deployable memory because the relation/index table must also be stored, limiting 32-bit floating-point (FP32) weight-memory reduction to about 4× under one-byte indexing. At matched parameter count, structured and dense models have comparable flash and instruction requirements, so structure improves Burgers accuracy and antisymmetry at a given footprint rather than automatically shrinking hardware requirements. Under complete lossless checkpointing, interruption timing has no independent effect, while budget loss and checkpoint degradation govern robustness. In the eleven-problem dataset, none of four tested descriptors reaches statistical significance for budget sensitivity. These results separate what physics structure preserves from what TinyML deployment must measure explicitly.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090266

Authors: Yerzhan Kenzhebek Timur Imankulov Samson Dawit Bekele Saltanbek Mukhambetzhanov

Physics-informed neural networks (PINNs) are difficult to train on strongly coupled, multi-objective systems: with fixed loss weights an otherwise identical run succeeds or diverges with the random seed, and a leading explanation attributes the failures to conflict between the loss-term gradients. We test that explanation on a coupled non-isothermal Buckley–Leverett displacement with a passive tracer, a three-front hot-water flood with an exact analytical solution as ground truth. The test produces an exactly verified counterexample: the conflict-free update ConFIG attains its alignment objective yet converges to a stable solution that violates the equations, while a quasi-second-order optimizer (SOAP) and per-term gradient surgery (PCGrad) train the network reliably with no loss-weight search. The most aligned update is thus the one that fails: gradient alignment is neither necessary nor sufficient for trainability, and the conflict is a diagnostic of the loss rather than its cause. The same formulation serves the forward, parametric, and inverse problems: it predicts the movable-oil recovery with a relative error of about 5%, returns the solution for any oil viscosity in a trained range from one network, and, from sparse near-injector monitoring, recovers the thermal front speed and one transport coefficient when the other is known. To our knowledge this is the first PINN solution of this coupled displacement, which is central to thermal oil recovery.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090265

Authors: Miguel A. Zamora-Ursulo Amira Flores Elias Manjarrez

Deep visual recognition can turn ordinary video into interpretable motor knowledge, yet coordination among the joints of a single body during social interaction remains largely unexplored. We present an interactive playback visualizer that couples markerless pose estimation with the cross-wavelet transform to quantify amplitude co-modulation between all joint pairs. Amplitude co-modulation proved anatomically structured: bilateral homologous pairs exceeded cross-limb and head–body pairs even among pairs sharing no keypoint, where correlated tracking error cannot produce it. Within-limb pairs also scored high but share keypoints, so their elevation is confounded with measurement error and not treated as established. Co-modulation concentrated at low postural frequencies and showed no systematic temporal trend; each profile remained temporally consistent within the session, and equivalence to a flat trend was not established. Vocal activity correlated with upper-limb movement at gesture frequencies. Each participant’s 78-dimensional profile was individually distinctive: split-half identification reached 28.6% (95% CI 19.3–40.1%) against 1.4% chance, demonstrating within-session identifiability rather than a cross-session trait. No sex differences were detected in overall or category-level co-modulation, and equivalence was not established for any measure; women exceeded men in the micro-movement band, the narrowest and most attenuated by preprocessing, so that difference is reported but not interpreted.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090264

Authors: Yangbin Tan Liping Mo Yu Yan

Hyperparameter optimization (HPO) aims to identify high-quality model configurations under a limited evaluation budget. To address mixed search spaces, sparse feedback, and low sample efficiency in reinforcement learning (RL)-based HPO, a Center-Guided Reinforcement Learning (CGRL) method is proposed. In CGRL, the policy output is reformulated from a configuration to be directly evaluated into a search center that defines a promising region, decoupling region-level guidance from exact configuration selection. A mixed candidate pool is generated around the center, and a promising candidate for real evaluation is selected by a Random Forest surrogate model. Meanwhile, a process-aware reward provides dense and informative feedback for policy learning. Experiments on 20 Yet Another Hyperparameter Optimization (YAHPO) Gym environments validate the effectiveness of CGRL. Compared with random search (RS), Tree-structured Parzen Estimator (TPE), Sequential Model-based Algorithm Configuration 3 (SMAC3), a Proximal Policy Optimization baseline (PPO-basic), Hyperparameter Optimization by Reinforcement Learning (Hyp-RL), and Q-Learning for Hyperparameter Optimization (HyperQ-Opt), CGRL achieves the best average rank of 1.800 in terms of the final best objective value, versus 6.000, 3.600, 2.200, 4.450, 6.350, and 3.600, respectively. For Low-Rank Adaptation (LoRA) HPO for relation extraction (RE) from ancient Chinese historical documents, CGRL improves Macro-F1 by 8.66%, 3.10%, 3.18%, and 5.13% on the validation set relative to RS, TPE, SMAC3, and PPO, respectively, and by 11.15%, 2.03%, 3.11%, and 9.59% on the test set. These results demonstrate the effectiveness of CGRL for limited-budget HPO and its applicability to practical RE tasks.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090263

Authors: Chulei Zhang Yuesong Wu Xuan Wu Yubin Xiao You Zhou

Recent studies have proposed post-processing strategies that iteratively reconstruct a partial segment of current solutions using Neural Combinatorial Optimization (NCO) solvers, thereby improving their performance on large-scale Vehicle Routing Problem (VRP) instances. However, the reconstruction subproblem is essentially a Shortest Hamiltonian Path Problem (SHPP) instance, which differs fundamentally from the original VRP variant on which the solver is trained. Consequently, the NCO solver may suffer from performance degradation during reconstruction due to limited generalization across problem variants. Moreover, relying solely on the solver may be insufficient to effectively identify and correct complex intersections or suboptimal topological structures. To address these limitations, we propose a post-processing strategy termed Hybrid Reconstruction Framework (HRC). Specifically, HRC first fine-tunes the NCO solver on SHPP instances and then exploits the enhanced solver to perform large-neighborhood random reconstruction. Subsequently, HRC conducts small-neighborhood reconstruction using 2-opt and kNN-DGR. The experimental results on both synthetic and real-world Traveling Salesman Problem and Capacitated Vehicle Routing Problem instances demonstrate that HRC substantially improves the performance of two representative NCO solvers on large-scale instances and achieves better overall performance than the state-of-the-art reconstruction strategy. Finally, ablation studies further validate the effectiveness of all proposed designs.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090262

Authors: Farida Abdoldina Azamat Serek Guido Cervone

Large ensembles of atmospheric dispersion simulations are necessary for uncertainty quantification, source attribution, and emergency response, however, physics-based models like HYSPLIT might be too computationally expensive for operational use. In addition, they have not fully benefited from recent advances in artificial-intelligence-based surrogate modelling, which offer the potential to reduce this computational burden. This study introduces the Dual-Head Mass-Consistent Plume Surrogate (DM-CPS), a machine-learning surrogate to promote the prediction of PM2.5 transport at a much higher speed without compromising the geometry and integrated pollutant mass of the plume. The proposed framework utilizes two types of complementary histogram gradient-boosting models: one trained in the logarithmic concentration domain to reconstruct the plume morphology, and one trained in the original concentration domain to estimate the total pollutant burden. The outputs of the two branches are then combined through a field-level mass-consistency scaling process that preserves the spatial structure of the plume while enforcing a physically consistent integrated pollutant mass. The model is trained on a GDAS-driven ensemble of 96 HYSPLIT simulations over Almaty, Kazakhstan, which includes 1.88 million grid-cell records and physics-informed wind-aligned features. DM-CPS achieves the highest operational dispersion skill (FAC2 = 0.57) with a significant reduction in geometric bias (MG: 2.75 → 1.68) and an improvement in the median integrated-mass ratio (0.13 → 0.23) while maintaining plume structure (SSIM ≈ 0.83) compared to random forest, multilayer perceptron, and conventional gradient-boosted baselines. The surrogate predicts an entire concentration field in about 35 ms, a factor up to five times faster than running HYSPLIT. The cross-regime evaluation also shows a significant gap of generalization in the presence of unseen, out-of-training meteorological data, again emphasizing the need for increased diversity of meteorological training data for operational use. The results show that physically informed surrogate modeling can significantly speed up the prediction of atmospheric dispersion, which is fundamental to run large ensembles.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090261

Authors: Tong Li Zhipeng Du Xinyu Hu Yuanyuan Zhang Gang Xiao

Translating natural language into graph query languages (NL2GQL) enables non-expert users to access graph databases, but supervised parsers depend on large annotated corpora and costly retraining whenever the schema evolves. Few-shot in-context learning offers a training-free alternative, yet its effectiveness hinges on demonstration selection, and conventional retrieval based on sentence-level matching and a single similarity signal often overlooks beneficial exemplars. We propose MGDR-SDR, a training-free approach that applies constituency parsing to derive three complementary representations of the utterance: a shallow abstraction preserving fine-grained lexical and syntactic detail, a deep abstraction exposing the intent-level syntactic skeleton, and a noun phrase list capturing concrete graph elements. These representations drive a multi-granular retrieval procedure that aggregates evidence across dense and sparse channels, yielding broader and more precise demonstration coverage. We evaluate MGDR-SDR on SPARQL generation over KQA Pro and Cypher generation over ZOGRASCOPE. Without any training, it reaches 86.58% Hits@1 on KQA Pro, surpassing strong supervised parsers and setting a new state-of-the-art among few-shot methods, and it transfers across GQL formalisms, attaining the best iid accuracy and a balanced generalization profile that surpasses all open-source few-shot baselines on ZOGRASCOPE. These results show that syntax-derived, multi-granular retrieval is an effective and generalizable strategy for few-shot NL2GQL semantic parsing.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090260

Authors: Joaquín Manuel Sánchez-Martín Juan Luis Bejarano-Luque Carolina Gijón Matías Toril Salvador Luna-Ramírez

Network densification will play a vital role in next-generation cellular networks, as it addresses coverage and capacity issues in the radio access domain. However, new cell deployments can negatively impact neighbor cells, requiring careful planning. Unfortunately, the complexity of advanced radio resource management schemes implemented in Beyond 5G and future 6G networks makes it challenging to predict the impact of radio deployments in the existing cells accurately. This work presents GRIMCELL, a novel data-driven method to estimate the impact of radio deployments (i.e., new carriers, sectors, or sites) on cell performance. For this purpose, GRIMCELL combines radio planning, configuration management and performance management data gathered in commercial vendor equipment. The core of GRIMCELL is a proven message passing graph neural network (XENet) to forecast the performance of both new and existing cells after the deployment. Assessment over data from a commercial LTE-A Pro network has shown that GRIMCELL provides high accuracy and flexibility across diverse deployment scenarios.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090259

Authors: María Pérez Andrés Bastidas-Fuertes Monserrate Intriago-Pazmiño Lenin G. Falconi Juan Benavides

Background: Preeclampsia remains a major cause of maternal and perinatal morbidity and mortality, and machine learning (ML) and artificial intelligence (AI) models have increasingly been proposed for risk prediction, diagnosis, monitoring, and prognosis. However, the extent to which key methodological details are visible from abstracts and bibliographic metadata remains unclear. Methods: We conducted an abstract-level systematic mapping review of ML/AI studies for preeclampsia (PE), hypertensive disorders of pregnancy (HDP), and directly related complications. Records were identified from ACM Digital Library, Elsevier/ScienceDirect, IEEE Xplore, PubMed, Scopus, SpringerLink, and snowballing. After DOI-based deduplication, abstract-availability checks, technical metadata filtering, and a two-stage eligibility audit, 29 primary ML/AI studies were retained for the final synthesis. Two authors independently re-audited the 145 technically retained records for eligibility. Initial agreement was 131/145 records (90.3%); for the binary include/exclude decision, Cohen’s kappa was approximately 0.764, indicating substantial agreement. Extraction was limited to title, abstract, year, DOI, and bibliographic metadata using a predefined RQ-aligned schema. Therefore, all frequencies represent information explicitly extractable at the abstract level and should be interpreted as lower-bound reporting estimates rather than full-text prevalence estimates. Results: Among the 29 eligible studies, tree-based models were the most frequently identifiable family (17/29), followed by linear/generalized linear models (15/29), support vector machines (9/29), and neural networks (8/29). The most frequently identifiable input modalities were EHR/clinical variables (21/29), demographics (10/29), laboratory variables (7/29), vital signs/blood pressure (7/29), and ultrasound/Doppler information (6/29). AUC was mentioned in 20/29 abstracts, but a numeric AUC value was extractable into the structured field in 14/29. These values were summarized descriptively and were not pooled because outcomes, prediction horizons, populations, predictors, and validation strategies were heterogeneous. External validation was explicitly identifiable in 8/29 abstracts, explainability-related methods in 5/29, and repository or dataset-source information in 1/29. Conclusions: The abstract-level evidence shows an active but heterogeneous ML/AI literature for PE/HDP prediction and related outcomes. However, abstracts often provide insufficient detail for robust comparative assessment of model validity, calibration, clinical readiness, reproducibility, and implementation. Future work should prioritize full-text evidence synthesis, standardized reporting, external and temporal validation, calibration, equity assessment, clinically meaningful explainability, and reusable data/model resources.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090258

Authors: Muntasir Adnan Manile Srun Carlos C. N. Kuhn

The original ALPHA benchmark introduced a taxonomy-aware penalty for evaluating CWE-level vulnerability prediction in Python and proposed that the penalty could theoretically also serve as a training signal. This paper tests that proposal empirically, comparing three delivery mechanisms–supervised fine-tuning, a dual-head classification loss, and reinforcement learning with a dense reward derived from the normalised penalty. We find that supervised approaches consistently regress below the zero-shot baseline under distribution shift, while GRPO succeeds and generalises across independent training seeds. Using a default, non-test-informed configuration, GRPO reduces the mean cumulative ALPHA penalty of Qwen2.5-Coder-7B-Instruct on Security Hardening and Adversarial Testing (SVEN) by 13.5%. A further test-selected configuration reduces the mean penalty by 22.9% (27.9% in the best-observed run), reaching statistical parity with its 4.5× larger zero-shot teacher. Our experiments show that reinforcement learning outperforms the alternative delivery mechanisms we test. We conclude that delivery mechanism, rather than the specific hierarchical structure of the reward, is the primary variable—though whether taxonomy-aware structure contributes further remains an open question.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090257

Authors: Ehsan Hallaji Roozbeh Razavi-Far

Continual learning has made significant progress toward enabling adaptive machine learning under evolving environments, yet real-world deployment increasingly exposes systems to persistent harsh conditions, including distributional shifts, feature evolution, delayed or scarce supervision, imbalance, noise, and recurring or novel classes. While prior research has largely addressed these challenges in isolation, growing environmental complexity motivates a broader rethinking of continual adaptation as a self-regulating process rather than solely a parameter update problem. Building upon the emerging framework of Self-Adaptive Learning (SAL), this perspective explores how learning systems may progress beyond reactive adaptation toward autonomous recognition, policy selection, and context-sensitive regulation of learning behavior under persistent uncertainty. Rather than proposing a specific algorithmic solution, we position SAL as a conceptual systems framework for organizing future research on resilient, long-lived machine learning systems. We discuss key implications for deployment robustness, evaluation, safety, and adaptive governance, while outlining major open challenges in developing practical self-regulating learners. By strengthening SAL as a forward-looking framework, this work aims to advance the broader conversation on machine learning systems capable of sustained autonomy in dynamic real-world environments.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090256

Authors: Andrea Moral Antonio Arroyo Juan Aparicio Xavier Barber

Clinical practice guidelines require expert synthesis that large language models (LLMs) might partly automate, yet their ability to reproduce clinically actionable recommendations is poorly quantified. We evaluate an LLM (Claude Sonnet 4.6) against the 103 recommendations of the Spanish enhanced-recovery guideline Vía RICA 2026, grouped in 17 bundles. The model used the panel’s own closed corpus (617 documents) in a multilingual retrievalaugmented generation pipeline. Concordance was assessed twice: by optimal 1:1 bipartite matching (Hungarian) on cosine similarity, and by an LLM-as-a-judge clinical adjudicator (Claude Haiku 4.5) validated against a three-clinician panel (Fleiss’ κ = 0.538). The two schemes bracket a micro F1 of 0.61–0.69 and reveal four findings: (i) a systematic granularity bias, producing 1–8 recommendations per bundle regardless of ground-truth size; (ii) failure of cosine similarity to discriminate within narrow clinical domains; (iii) high reference-concordance precision (0.70–0.81) despite low exhaustiveness; and (iv) no transfer of the GRADE fields, evidence level agreeing no better than chance and strength systematically downgraded. An eight-fold larger retrieval budget left it intact. A corpus audit found 25 documents that formulate recommendations; excluding them lowers judged micro F1 to 0.602. The results delimit the current utility of generative AI for guideline development.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090255

Authors: Rojaina Mahmoud Omneya Attallah Ahmad Al-Kabbany

We investigate the potential of technology-based attention training (AT), particularly virtual reality (VR), as a stress-management tool. Mental stress is rising globally, and researchers increasingly use immersive technologies, wearable sensors, and machine learning (ML) for its detection and control. This feasibility study examines the impact of fully immersive VR-based AT on mental stress using electroencephalogram (EEG) signals and automated classification. We designed virtual exercises targeting different attention types and analyzed EEG responses with an ML framework; the resulting dataset, collected at the Arab Academy for Science and Technology (Alexandria, Egypt), is publicly available. For an unbiased estimate, we adopt a leakage-free evaluation in which the train/test split is performed by time, before segmentation into overlapping windows, so neighboring windows cannot appear in both sets. Subject-specific tree-based classifiers detected stress with a mean accuracy of about 97%, whereas leave-one-subject-out (LOSO) validation yielded about 67%, indicating strongly individual stress signatures and motivating a subject-specific strategy. Using these models, we compared the number of classifier-predicted stress segments before and after AT and visualized the feature space with T-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP). Under subject-specific models the number of stress-predicted segments decreased after AT (Wilcoxon signed-rank p<0.05); however, because this reduction was corroborated neither by a non-circular (LOSO) detector nor by a centroid-separation measure, we interpret it as an exploratory, classifier-predicted effect rather than independently validated stress reduction. The results highlight the promise—and the current limits—of integrating immersive VR with EEG-guided analytics for mental-health support.

Machine Learning and Knowledge Extraction doi: 10.3390/make8090254

Authors: Melike Tombaz Andreas K. Nüssler Engin Tercan Niklas R. Braun Andreas Fritsche Tina Histing Nico Pfeifer Sabrina Ehnert

Undiagnosed diabetes and prediabetes are common among trauma patients and increase perioperative complication rates, yet routine glycemic screening at hospital admission is rarely performed. We compared 10 supervised machine learning (ML) algorithms for early detection of impaired glucose metabolism in 1789 trauma surgery patients (1095 normal; 694 prediabetes/diabetes) using questionnaire data and preoperative laboratory values within a nested cross-validation framework with hyperparameter optimization. On the combined feature set, XGBoost achieved the highest area under the receiver operating characteristic curve (ROC-AUC; 0.815), followed by multilayer perceptron (0.814) and Gradient Boosting (0.813); adding blood parameters significantly improved performance for nine of ten algorithms (DeLong’s test, p < 0.05). SHapley Additive exPlanations (SHAP) and XGBoost feature gains identified six overlapping predictors, yielding a simplified model with comparable mean discrimination. Inter-model differences were small, and a fully interpretable logistic regression model remained within one percentage point of XGBoost. The best-performing model demonstrated strong calibration and a higher net benefit than either universal testing or no testing across the evaluated threshold probabilities. In routine trauma practice, such a model could identify patients who would benefit from confirmatory glycated hemoglobin (HbA1c) testing, enabling earlier detection of dysglycemia, risk-stratified glucose management, and timely referral for lifestyle intervention in patients with prediabetes.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080253

Authors: Geofrey Owino Timothy Kamanu John Ndiritu

Adaptive decision-making has emerged as an important direction in audio classification. Reinforcement learning (RL) and multi-armed bandit (MAB) methods offer principled frameworks for sequential and localized decision-making. However, despite growing interest in these approaches and the availability of review studies on audio classification and audio-based learning, their role within the audio classification pipeline remains fragmented and has not been systematically synthesized. This study presents a systematic review of RL- and MAB-based approaches in audio classification, to characterize how adaptive decision-making is formulated, where it is applied within the pipeline, and what impact it has on system performance, robustness, and efficiency. The review followed PRISMA guidelines. A literature search across IEEE Xplore, Scopus, Nature, and Google Scholar identified 1896 records. Thirty-one studies met the predefined inclusion criteria and data were extracted for analysis. Reporting quality was assessed using the TRIPOD framework, and methodological reliability was evaluated using a domain-specific risk-of-bias tool. The findings show that adaptive decision-making in audio classification is evolving toward a control-centric paradigm. RL-based optimization and control approaches dominated the literature, whereas bandit-based approaches remained comparatively underused. A central finding of the review is a structural mismatch between problem type and method choice. Many adaptive tasks are inherently local and repeated decision problems, yet they are predominantly addressed using full RL frameworks rather than lighter bandit formulations. This review introduces a taxonomy of adaptive decision-making mechanisms and proposes a unified framework that reframes audio classification as a layered decision-making process under uncertainty. The evidence suggests that the future of audio classification lies not only in improved prediction, but in adaptive decision-making architectures capable of managing uncertainty, variability, and real-world deployment constraints.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080252

Authors: Catalin Anghel Andreea Alexandra Anghel Marian Viorel Craciun Adina Cocu Simona Moldovanu Cristian Sandu Christiana Diana Maria Dragosloveanu Constantin Adrian Andrei Diana-Elena Vulpe Calina Maier Cristian Scheau Serban Dragosloveanu Vasile Potop

Background: Large language models (LLMs) are increasingly explored for automated educational assessment, while future educational platforms may combine grading, feedback, learner analytics, and personalization. The objective of this study was to determine whether student-history metadata can influence the numerical score assigned to the same answer. Methods: This study introduces GradeDrift-LLM, a controlled framework for measuring student-history-induced score drift in LLM-based automated grading. We evaluated 1000 Computer Science answers from 100 students across six student-history conditions and eight open-weight LLMs. For each grading instance, the submitted answer, question, reference answer, rubric-related information, scoring scale, and grading instruction were kept constant; only the student-history condition varied. Results: Across 39,997 valid paired comparisons, 83.92% showed no drift, 9.40% showed upward drift, and 6.68% showed downward drift. Mean absolute drift was 0.2137 points, and the 95th percentile absolute drift was 1 point. Positive-history frames tended to increase scores, whereas negative-history frames tended to decrease them. Drift was model-dependent, not uniformly explained by approximate scale, and present in both technical and argumentative answers; rare extreme deviations reached 10 points. Conclusions: Student-history metadata can influence LLM-generated grading scores despite explicit instructions to ignore it. Future LLM-based grading systems should separate answer-based scoring from learner-context-based personalization and validate score invariance under controlled learner-context variations.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080251

Authors: Patrik Viktor Gabor Kiss

Autonomous vehicles increasingly require capabilities that extend beyond perception towards contextual understanding, semantic reasoning, and explainable decision-making. Knowledge graphs (KGs) have emerged as a promising solution by integrating heterogeneous sensor data, traffic regulations, domain knowledge, and contextual information into unified semantic frameworks. This review systematically examines knowledge graph-based intelligent reasoning in autonomous driving through a PRISMA 2020-guided analysis of 47 peer-reviewed studies identified from the literature published from 1 January 2018 to 31 January 2026. The findings reveal that semantic scene understanding and ontology-based representations currently dominate the field, with 66.0% of studies integrating knowledge graphs with deep learning approaches. Neuro-symbolic methods and explainable AI components were identified in 38.3% and 34.0% of publications, respectively, indicating increasing research interest in hybrid and transparent AI architectures. The analysis further demonstrates that 80.9% of studies remain limited to benchmark datasets and simulation environments, whereas only 19.1% provide real-world validation, suggesting relatively low technological maturity and limited industrial readiness. Although KG-enabled approaches substantially improve contextual awareness, hidden hazard anticipation, and explainability compared with conventional perception-centric architectures, major challenges remain regarding scalability, ontology interoperability, semantic error propagation, real-time reasoning, and certification requirements. The review identifies the convergence of knowledge graphs, large language models, and neuro-symbolic AI as a promising direction for next-generation autonomous driving systems. Future research should therefore focus on uncertainty-aware reasoning, adaptive explainability, standardised evaluation methodologies, and certification-oriented real-world deployment strategies.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080250

Authors: Batu Kaan Özen Thomas Waschulzik Alois Knoll

Machine learning (ML) models have achieved remarkable success in areas ranging from healthcare to autonomous systems. Yet, their inherent complexity frequently obscures the reasoning behind their decisions, undermining transparency, accountability, and user trust. Compounding this issue is the absence of a universal methodology for comparing and ranking interpretability across diverse ML models and interpretability techniques. This paper addresses this gap by introducing a configurable framework of quantitative metrics to evaluate and rank interpretability. Our approach offers a structured, heuristic basis for assessing model clarity, decision logic, and accessibility, enabling practitioners to systematically compare interpretability across a wide range of algorithms and techniques. The resulting scores are intended as a practical, domain-configurable heuristic guide for comparison rather than a universal notion of interpretability.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080249

Authors: Jiehua Qi Haoran Wang Rui Wang Xinxiao Wu Hanhong Hu Bingcong Chen

Remaining useful life (RUL) estimation is a fundamental task in Prognostics and Health Management (PHM), supporting condition-based and predictive maintenance of engineering systems. Data-driven methods contribute to many effective strategies for RUL prediction. However, two problems need to be solved when they are used in industrial applications: (1) The amount of data under one working condition is limited, and data from different working conditions suffer from domain discrepancies. These methods are constrained by distribution differences in data under different working conditions. (2) There is an urgent need to quickly achieve prediction with much less computing resources. To address these issues, a lightweight RUL prediction method called a knowledge transfer-based heterogeneous distillation network is proposed by combining knowledge distillation and transfer learning. First, the adversarial training mechanism is introduced for the extraction of domain-invariant features. Subsequently, a heterogeneous knowledge distillation framework is further designed for remaining useful life prediction, in which a bi-directional long short-term memory model serves as the teacher network and a compact fully connected network acts as the student model. The teacher model is used to learn informative degradation patterns and guide the training of the lightweight student model through knowledge transfer. Results obtained on the N-CMAPSS dataset verify that the proposed method achieves promising effectiveness and strong generalizability, reducing the average RMSE and MAE by 44.83% and 41.30%, respectively.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080248

Authors: Alban Bornet Abiram Sandralegar Anthony Yazdani Elias Adam Benguettat Michael Francis Righini Feres Ravarelli Paul Eugène Constanthin Alexandre Lavé Julien Haemmerli Insa Janssen Jelia Issa Elham Qaderdan Ethan Guillaume Godin Nasser Bouhalassa Karl Schaller Philippe Bijlenga Douglas Teodoro

This study conducts a clinical evaluation of a secure, locally deployed, quantized large language model (LLM) for automating modified Rankin Scale (mRS) score extraction from unstructured neurosurgical notes. We retrospectively selected 103 authentic clinical letters (2007–2025) from aneurysm patients at a tertiary neurosurgical centre. To comply with data privacy constraints, an open-source reasoning LLM (Qwen3-32B) with 4-bit quantization was deployed entirely on-premises. The LLM extracted mRS scores using a zero-shot approach with custom logits processors to enforce strict JSON formatting. Performance was compared to a reference standard (attending neurosurgeons’ consensus) and parallel scoring by medical residents and students. The LLM achieved excellent agreement with the attending consensus (QWK 0.95), matching the reliability of medical students (QWK 0.95) and residents (QWK 0.93). Exact agreement was 75%, and agreement within ±1 mRS point was 96%. Bayesian analysis strongly supported statistical equivalence between the model and human raters. The computationally optimized LLM demonstrated human-level classification reliability without task-specific fine-tuning. This approach successfully addresses key patient data privacy barriers and the formatting inconsistencies typical of open-ended generative models. Securely deploying a general-purpose, quantized LLM provides a scalable pathway for extracting functional outcomes and supports FAIR-aligned data systems.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080247

Authors: Jiao Tian Jinlin He Liejun Wang

Client selection is a critical mechanism for ensuring robust convergence in Federated Learning (FL) systems, yet it remains vulnerable to Non-IID data distributions and Byzantine attacks. Deep Reinforcement Learning (DRL) has shown promise for automated client selection, yet existing methods suffer from three structural deficiencies: observation ambiguity, where scalar states cannot distinguish malicious updates from benign heterogeneity; reward fragility, whereby attackers exploit unbounded feedback to hijack policy updates; and risk blindness, as risk-neutral agents overlook the inherent variance in client contributions. To address these deficiencies, we propose InfRA-FL, a robust adaptive framework. A mutual-information-based state construction extracts high-utility features, resolving observation ambiguity. A Saturation-Aware Robust Reward (SARR) mechanism applies soft-clipping to bound each client’s influence on the policy gradient, provably neutralizing reward poisoning. URA-PPO, an uncertainty-aware algorithm with a dual-head Critic, optimizes a risk-penalized objective that shifts the agent from risk-neutral to risk-averse decision-making. Experiments on MNIST and CIFAR-10 under 20% Byzantine adversaries show that InfRA-FL outperforms state-of-the-art baselines by 5–10% in accuracy while accelerating convergence, establishing that principled information-theoretic observation and robust reward design suffice to secure RL-driven federated learning against targeted poisoning.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080246

Authors: Rakhat Meiramov Tomiris Rakhimzhanova Adil Taibassarov Zhanat Makhataeva Huseyin Atakan Varol

Target-speaker automatic speech recognition (TS-ASR) enables transcription of a specific speaker in multi-talker environments, yet remains largely unexplored for multilingual, low-resource languages. Existing TS-ASR systems predominantly target monolingual English using diarization-based or speaker-embedding approaches, leaving a critical gap for languages such as Kazakh, where code-switching with Russian and English is commonplace. We propose Persona-ASR, a modular two-stage architecture. The first stage is an explicit target-presence gate that verifies whether the enrolled speaker appears in the mixture and emits a <no_target> token to suppress transcription when the speaker is absent, directly addressing the acoustic-hallucination failure mode of prior systems. The second stage performs enrollment-conditioned recognition: a 192-dimensional ECAPA-TDNN speaker embedding modulates a WavLM-Base-Plus encoder through feature-wise linear modulation (FiLM), while language-specific CTC heads enable joint Kazakh and English decoding without forcing Latin and Cyrillic symbols to compete in a single output space. To evaluate the system, we introduce KazMix3, a Kazakh overlap dataset for TS-ASR training, and PersonaMix, a controlled bilingual benchmark spanning same- and cross-language enrollment across varying interferer counts (1–3) and signal-to-noise ratios (−3 to +3 dB). Persona-ASR outperforms a strong off-the-shelf cascade baseline by 13.3 WER points on English and 24.6 on Kazakh, and matches a published monolingual English baseline. On PersonaMix, speaker conditioning reduces relative word error rate by 40.7% on English and 59.3% on Kazakh mixtures over an unconditioned variant of the same model, and cross-language enrollment (unseen during training) remains effective, increasing average raw WER by only 4.1 points (English) and 2.2 points (Kazakh) relative to same-language enrollment. To our knowledge, Persona-ASR is the first TS-ASR system for the Kazakh language, establishing a foundation for multilingual personalized ASR in low-resource settings.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080245

Authors: Shymaa Darwish Mohamed Mohamed El-Habrouk Ayman Samy Abdel-Khalik Ragi Ali Rifaat Hamdy

Reliable digital twins of complex nonlinear systems require not only high predictive accuracy but also physical consistency, robustness under degraded operating conditions, and explicit uncertainty handling. Purely data-driven models often suffer from poor generalization, unphysical behaviors, and limited interpretability when facing noisy measurements and unseen operating conditions. This paper introduces a knowledge-guided physics-informed hybrid learning framework that integrates recurrent neural networks with Unscented Kalman Filter (UKF) state estimation and embedded thermodynamic constraints within a unified uncertainty-aware architecture. The proposed PI-LSTM-UKF framework achieves competitive predictive accuracy and improved physical consistency relative to the residual-learning hybrids by tightly integrating physics-informed recurrent learning, thermodynamic constraints, and sequential UKF state estimation. While the UKF provides robust recursive correction under noisy measurements during closed-loop operation, the physics-informed Long Short-Term Memory (PI-LSTM) learns nonlinear corrections and long-term dynamics that cannot be captured by the linear model alone. The proposed framework is systematically benchmarked against a hierarchy of seven modeling approaches, including Dynamic Mode Decomposition with control (DMDc), Sparse Identification of Nonlinear Dynamics (SINDy), and residual-learning variants based on Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). High-fidelity Simscape simulations of a Rankine-cycle steam turbine system are used as a challenging simulation-based case study. Results show that the knowledge-guided hybrid approach achieves competitive predictive accuracy, improved physical consistency, and robust performance under an unseen load profile, severe thermodynamic degradation, valve hysteresis, and substantially elevated sensor noise. The framework provides a promising simulation-based foundation for uncertainty-aware digital twins of nonlinear thermal power systems. Validation using operational plant data remains necessary before its application to real-time monitoring and predictive maintenance.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080244

Authors: Sergei O. Kurashkin Vadim S. Tynchenko Aleksei S. Borodulin Vladimir A. Nelyub Nikolay O. Kalutsky Tee Connie

Tabular foundation models (TFMs) have moved tabular machine learning from per-dataset training towards amortised in-context inference, fitting a small-to-medium table in a single forward pass without a training run. The 2024–2026 release train, the TabPFN and TabICL lines and challengers such as Mitra, LimiX and Orion, has produced a generation whose architectures, capabilities and limits are documented mainly in preprints, while existing surveys treat these models as a subsection of tabular deep learning or of language-model table understanding. This review is, to our knowledge, the first organised around the current generation. From a corpus of 961 screened records and 98 retained studies, it taxonomises the architectures by pretraining regime, maps the capability space across five axes, isolates the language-model-on-tabular strand for prediction, feature engineering and generation, and summarises openness and deployment. A dedicated critical synthesis then reads the reported capabilities against independent evidence: on the studies reviewed here, tree-based and deep models retain the lead across 142 curated datasets that go beyond the standard independent and identically distributed setting; on 112 datasets, the models attain the highest accuracy but weaker conditional coverage than gradient-boosted trees; and robustness under feature shift, fairness and generation quality remain open. Amortised in-context prediction is thus a working paradigm whose independent evidence has yet to match its benchmark claims.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080243

Authors: Gregory Gorbov Artem Latyshev Aleksandr Panov

This work investigates the safe exploration problem in reinforcement learning, where an agent must maximize cumulative performance while simultaneously satisfying safety constraints. This challenge becomes particularly difficult in long-horizon tasks because safe reinforcement learning methods often become overly conservative, restricting exploration and preventing agents from reaching distant goals, while estimation errors accumulate over extended horizons. We propose Imagine To Ensure Safety in Hierarchical Reinforcement Learning (ITES), which combines a learnable world model with high-level and low-level policies to promote safety at both hierarchical levels. The novelty of ITES lies in jointly ensuring safe subgoal generation and safe subgoal execution within a hierarchical framework. The high-level policy generates intermediate subgoals that guide exploration toward safe regions, while the low-level policy uses imagined rollouts in the learned world model to reduce unsafe behavior during subgoal execution. We evaluate ITES on the long-horizon SafeAntMaze C-shape, SafeAntMaze W-shape, and SafePusher benchmarks, as well as on short-horizon Safety Gym tasks, using task performance and episodic cost under prescribed safety budgets. The results show that ITES achieves substantially higher task success and constraint-compliant success rates on the long-horizon benchmarks while maintaining mean episodic costs below the prescribed budgets.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080242

Authors: Khairil Imran Ghauth Yanche Ari Kustiawan

Quantum machine learning (QML) has emerged as a promising approach for analyzing the complex, high-dimensional data encountered in oncology, yet research in this area remains fragmented. This systematic literature review synthesizes current applications of QML in cancer care. Following PRISMA guidelines, six databases were searched for peer-reviewed English-language studies published between 2020 and 2026. Of the 212 records identified, 49 studies met the inclusion criteria after screening and quality assessment. The findings show that QML research is dominated by classification and detection tasks, while segmentation is beginning to emerge. Breast cancer and brain tumors are the most frequently investigated domains. Hybrid quantum-classical models, particularly quantum kernel methods, quantum neural networks, and quantum convolutional neural networks, are the predominant approaches. The main barriers to adoption are hardware limitations, including quantum noise, limited qubit availability, and reliance on simulators. Overall, QML in oncology remains in its early stages of development, with limited clinical validation, insufficient model interpretability, and little evidence of a clear quantum advantage. Future research should prioritize evaluation on real quantum hardware, larger and more diverse clinical datasets, standardized benchmarking against classical methods, and closer collaboration between computer scientists and oncology experts to facilitate clinical translation.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080241

Authors: Eduard Barnoviciu Corneliu Florea

Dataset distillation has achieved strong results in computer vision, but is largely underexplored in the tabular domain. We introduce a tabular dataset distillation method that projects the data through many random embedders (views) to achieve invariance to transformation and to focus on the consistency between real and synthetic sets. The synthetic set is determined through a formulation of distribution matching between the many-view projection of the original and distilled dataset. Building on this approach, our proposal achieves three goals: (1) we formulate the Tabular Alignment via Moment Embeddings (TAME) method and, by extensive empirical evaluation, we show its efficiency; (2) we evaluate TAME on a benchmark of 18 tabular datasets, with strong baselines and evaluation metrics; and (3) we present a structured set of studies analyzing the impact of losses, dataset geometry, embedder architecture, instances per class (IPC) budget and downstream classifiers. We show that the proposed TAME method consistently surpasses baselines on neural classifiers, while remaining competitive with strong coreset baselines on tree-based classifiers (RF, XGBoost). Performance is further increased, especially for tree-based classifiers with a lightweight validation method. Extensive evaluation, including on additional large-scale sets and ablation experiments, allow a better understanding of the method.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080240

Authors: Angel de Jesus Castro-Romero Julio Cesar Ramos-Fernández Marco Antonio Márquez-Vera Juan Manuel Xicoténcatl-Peréz Salatiel Garcia Nava Jorge Alberto Ruiz-Vanoye Sébastien Paris

Autonomous agricultural robots require accurate trajectory tracking to perform precision field operations such as seeding, fertilization, and pruning. Classical kinematic models fail to capture the nonlinear dynamics inherent to real field conditions, limiting the performance of model-based control strategies. This work proposes an optimal hybrid methodology integrating Takagi–Sugeno (T–S) fuzzy model identification and Pure Pursuit (PP) control within a Particle Swarm Optimization (PSO) framework for a simulated pruning tractor. Data-driven T–S fuzzy models for incremental displacements MΔx and MΔy are identified using Fuzzy C-Means and parameterized via PSO. These fuzzy models are embedded in a PP feedback control scheme with discrete-time PI velocity and PD steering controllers, whose four gains are tuned by a second PSO instance. The fuzzy models achieve identification Root-Mean-Square Errors (RMSEs) of 10.598 × 10−3 m and 8.125 × 10−3 m. Integrated into the control loop, the system yields a lateral RMSE of 6.6 × 10−3 m on the training path and generalizes effectively across twelve complex agricultural coverage trajectories, maintaining a lateral RMSE below 12 × 10−3 m and heading RMSE under 1 degree. This interpretable, fuzzy rule-based approach provides an accurate and replicable simulation baseline for future experimental implementation on physical platforms.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080239

Authors: Samuel Kurian Roy Sreehari M. S. Mohammed Fayyas N.M. Ekaterina Kopets Denis Butusov Sishu Shankar Muni

This research introduces a new spectral pruning approach using the Adjacency Edge Index (AEI), which is a centrality measure from the theory of spectral graphs and first-order matrix perturbation theory. The AEI score reflects the contribution of each neuron to the network’s dynamic synchronizability via the Fiedler vector. The AEI approach thus offers a mathematically motivated saliency score in the context of data-driven neuron co-activation graphs. The proposed approach has been tested on MNIST, Fashion MNIST, KMNIST, and a real-world social network dataset. The approach has been extended to convolutional filter pruning on the CIFAR-10 dataset using spatial global average pooling. The AEI approach has been compared to magnitude-based pruning methods like L1 and L2 norms and gradient-based pruning methods like SNIP and GraSP. The robustness of the proposed approach has been established by comparing the results over five random seeds. The AEI approach is proposed as a principled, interpretable, structure-aware pruning criterion rather than an accuracy-maximising method. Spectral analysis demonstrates that AEI is the only evaluated method that systematically targets structurally peripheral neurons, whereas magnitude-based methods prune broadly across the structural spectrum and GraSP actively removes structurally central neurons, an effect most pronounced on more complex datasets and deeper architectures (CIFAR-100, ResNet-20), where it causes substantial accuracy degradation at high sparsity. The AEI approach has been extended to the Hybrid approach by combining the AEI and L2 norms. The Hybrid approach has been seen to improve the accuracy gap at 40% sparsity from 5.46 to 2.26 percentage points over the state of the art. The robustness of the proposed approach has been established by conducting ablation studies on the robustness of the approach to the selection of the graph. The approach has been seen to be moderately robust to the selection of the graph with Spearman’s ρ ≈ 0.70. The accuracy gap has been seen to be less than 0.2%.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080238

Authors: Aleksandar Ćirić Zoran Ristić Tamara Gavrilović Anđela Rajčić Snežana Đurković Željka Antić Miroslav D. Dramićanin

Luminescence thermometry has evolved through decades of research focused on optimising materials and on extracting temperature information from isolated spectral features such as luminescence intensity ratios, bandwidth, line shift and excited-state lifetime. Despite extensive material development, these conventional methods remain fundamentally limited by construction: only a small subset of pre-selected spectral features is exploited, while the bulk of the temperature-relevant information encoded in the full spectrum is systematically discarded. A paradigm shift is presented here: Luminescence Thermometry 2.0 (LT 2.0), implemented through the newly developed SpectraSensML platform, in which machine learning regression operates on the entire emission spectrum to deliver temperature readout. The approach is demonstrated on a Yb3+-doped phosphor emitting in the near-infrared biological transparency window across 100 to 700 K. Yb3+ is a particularly demanding case: only the single 2F5/2 multiplet emits, and its weakly thermally coupled Stark sub-levels yield modest sensitivity under conventional intensity-ratio thermometry. A total of 27 regression algorithms drawn from four families, namely tree ensembles, physics-aware regression models, kernel and instance methods, and neural networks, are systematically benchmarked. A sensor-fusion estimator that combines the first three principal components reaches an average root-mean-square error of 0.36 K on an unseen-temperature test set, a seven-fold improvement over the best luminescence intensity ratio variant. Standard normal variate (SNV) normalisation is identified as the most effective preprocessing strategy because it isolates the band-shape deformations that encode temperature. Single-component approaches that rely on the first principal component alone are shown to be quantitatively sub-optimal: multi-component regressors that exploit the first three principal components reduce the temperature uncertainty by close to an order of magnitude. The structural reason behind the failure of decision-tree ensembles on unseen temperatures is explained: their piecewise-constant predictions cannot interpolate beyond training set-points. The open-source SpectraSensML application used to obtain the results is released alongside the manuscript to enable reproducible community benchmarks.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080237

Authors: Ayman Diyab Ahmad Diyab Ishaan Dhillon

The increasing digitization of healthcare records and growing reliance on interconnected systems have increased hospitals’ exposure to cyber threats, including ransomware, brute-force attacks, and Structured Query Language (SQL) injection. Traditional cybersecurity approaches alone are often insufficient to address these threats, prompting growing interest in artificial intelligence (AI)-based decision-support tools. This paper evaluates the potential of ChatGPT for hospital cybersecurity and incident response while introducing a structured qualitative framework for evaluating Large Language Model (LLM)-generated cybersecurity recommendations in healthcare. Through three progressively designed experiments and a ransomware case study, we evaluate ChatGPT’s role in developing a hospital cybersecurity plan, detecting brute-force login attempts, responding to an SQL injection attack, and managing a ransomware incident. Responses are assessed using five evaluation dimensions: specificity, completeness, technical correctness, feasibility, and alignment with the National Institute of Standards and Technology (NIST) Cybersecurity Framework (CSF), including both explicit mapping and function coverage. The results demonstrate that ChatGPT provides structured, context-aware guidance that aligns well with NIST CSF 2.0 and addresses governance and third-party risks. However, the recommendations also exhibit limitations, including limited operational depth, assumptions about technology and regulatory environments, lack of prioritization for resource-constrained settings, and limited consideration of implementation costs. Overall, the proposed evaluation framework provides a systematic approach for assessing LLM-generated cybersecurity guidance, while the findings indicate that ChatGPT can serve as a valuable decision-support tool that should complement, rather than replace, qualified cybersecurity professionals.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080236

Authors: Aleksei Samarin Artem Nazarenko Egor Kotenko Aleksei Toropov Alexander Savelev Alexander Motyko Valentin Malykh

This paper presents a novel method for pruning deep neural networks based on the concept of flow, derived from the continuous modeling of signal propagation across layers. We derive flow functions for fully connected, convolutional, and self-attention architectures, and we propose a new iterative pruning algorithm, Iterative Flow-Aware Pruning (IFAP), that leverages these measures to identify and eliminate non-essential parameters while preserving critical information pathways. Extensive experiments across ten prominent architectures (including CNNs, vision transformers, and efficient mobile networks) on ten benchmark datasets demonstrate consistent accuracy–compression trade-offs: 81% of the evaluated configurations achieve a 60–81% reduction in computational cost relative to the corresponding baseline model. Furthermore, 97% of the evaluated configurations retain more than 98% of their baseline Top-1 accuracy. These results validate flow-based importance scoring as a robust and general-purpose foundation for model optimization.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080235

Authors: Mark Korang Yeboah Nana Yaw Asiedu

Temperature–time profiles obtained through thermistor-based monitoring provide a rich but noise-sensitive source of information for estimating kinetic and thermal parameters in exothermic batch reactions. Conventional workflows typically combine deterministic smoothing with numerical differentiation, an approach that can amplify measurement noise and fail to propagate preprocessing uncertainty into the resulting reaction-rate and parameter estimates. To address these limitations, this study presents an uncertainty-aware Bayesian machine-learning framework that integrates scalable random-Fourier-feature Gaussian-process (RFF–GP) smoothing, analytical differentiation, temperature-derived apparent conversion, Bayesian parameter inference, posterior validation, predictive calibration, model comparison, ablation, sensitivity analysis, probabilistic benchmarking, simulation of thermal nonideality, and endpoint diagnostics. The framework was applied to 379,631 cleaned thermistor observations. The production RFF–GP achieved a validation root-mean-square error of 0.04805K, yielding a stable latent temperature trajectory and an uncertainty-aware estimate of dT/dt. On a smaller matched subset, exact Gaussian-process regression achieved the highest predictive accuracy and the best probabilistic scores, whereas the RFF–GP reduced central-processing-unit runtime by approximately 4.1-fold and remained applicable to the larger production fit. A Monte Carlo dropout neural comparator produced larger prediction errors and substantially wider predictive intervals. Six apparent thermokinetic structures were evaluated using mean-field variational inference, after which the nth-order and autocatalytic structures were validated using the No-U-Turn Sampler (NUTS). Under mean-field variational inference, the apparent autocatalytic structure achieved the lowest point estimate of the widely applicable information criterion (WAIC), the lowest derivative-domain error, and the lowest full-profile temperature-reconstruction root-mean-square error of 0.2920K. Its posterior obtained using NUTS yielded Ea=40.98kJmol−1, kref=0.005815min−1, ΔTad=56.11K, m=0.1694, and n=1.0784. The sampling diagnostics indicated satisfactory convergence, large effective sample sizes, and no divergent transitions. Although the MFVI posterior means and NUTS posterior medians were similar, variational inference produced narrower uncertainty intervals for several correlated parameters. Moving-block bootstrap intervals did not establish a decisive separation in WAIC among the leading structures. Expanded sensitivity, ablation, imperfect-insulation simulation, and endpoint-holdout analyses further showed that the apparent parameter estimates were sensitive to optimization, thermal nonideality, sensor response, and Gaussian-process boundary behavior. The autocatalytic formulation should therefore be interpreted as the best-performing apparent structure among the candidates tested rather than as evidence of a unique chemical mechanism. Overall, the framework extracted physically plausible apparent thermokinetic information from noisy temperature-only measurements while explicitly quantifying uncertainty arising from prediction, parameter estimation, model form, computation, thermal nonideality, and boundary behavior.

Machine Learning and Knowledge Extraction doi: 10.3390/make8080234

Authors: Héctor R. Becerril Villamil Vladimir Rodriguez Perez Daniel Sanin-Villa Julio Antonio Caballero-Mora Juan C. Tejada

Selective algorithmic auditing requires deciding which process evidence should receive attention when exhaustive review is infeasible. This Review introduces a four-layer framework that connects process representation, learning, inspection allocation, and governance within a single budgeted sequential decision problem over event streams. Unlike prior reviews centered on predictive process monitoring, explainability, cost analysis, or bibliometric structure, the proposed framework examines how these functions interact when human review, computation, latency, and documentation capacity are constrained. A structured and targeted survey of 89 unique publication families is used to illustrate and critically examine event-log, Petri-net, graph, object-centric, neural, uncertainty-aware, sequential, bandit, reinforcement learning, and audit architecture approaches. The reviewed evidence indicates that substantial bodies of work address the individual layers, but cross-layer evaluation remains fragmented and uses heterogeneous datasets, objectives, and validation protocols. The synthesis identifies five priorities: audit-ready benchmarks, explicit inspection budget protocols, calibrated uncertainty, transfer across organizational contexts, and reproducible governance interfaces. The main contribution is a computational framework and a corpus-bounded research agenda that connects representation, learning, inspection allocation, and governance for selective algorithmic auditing.