Mach. Learn. Knowl. Extr., Volume 8, Issue 7 (July 2026) – 49 articles

Memory-bank methods such as PatchCore are widely used in industrial quality control for visual anomaly detection since they require no training, are fast to deploy and achieve strong accuracy. However, they are memory-intensive. Furthermore, a single production line typically involves different products or cameras, so using a single anomaly detection method with a shared nearest-neighbor memory bank is attractive since it simplifies deployment and makes new products easy to add. Nevertheless, embeddings from different products/cameras can interfere during retrieval, causing what we call “memory pollution”. In this work, we study this effect through a new diagnostic framework, which involves: (1) a new metric, the wrong-neighbor rate (WNR), which measures how often a query’s nearest neighbor belongs to a different product; (2) an empirically validated phenomenon, “oracle inversion”, where querying only the product’s own data can underperform the shared bank under a fixed memory budget; (3) a first-order analytical model of the WNR, which predicts how pollution grows with product count and memory budget; and (4) a minimal training-free router that removes the effect of memory pollution. Our results show that our system performs robustly across different datasets and backbones, with up to 25× memory reduction, which makes our framework attractive for industrial applications. Full article

Validating wireless power transfer (WPT) systems for electric vehicles (EVs) is a challenge due to efficiency variations caused by coil misalignments and height differences arising from various vehicle designs. Traditional simulation methods, such as finite element analysis (FEM), provide high accuracy but entail significant computational costs and calculation times, limiting the number of case studies and their optimization. This paper presents a methodology that integrates Machine Learning (ML) and Genetic Algorithms (GA) to overcome these limitations. An ML model rapidly and accurately predicts key electromagnetic parameters across a wide range of positions and frequencies. These predictions feed into a GA that optimizes control variables (voltages and frequency) with the objective of maximizing power transfer efficiency, while simultaneously ensuring component integrity at each operating point. Beyond drastically reducing simulation time and experimental effort, this methodology will enable knowledge extraction and its use for formulating design rules. These rules can lay the groundwork for developing simplified, real-time adaptive control strategies, facilitating the reduction of control variables and the narrowing of search ranges. Full article

Alzheimer’s disease (AD) is the most prevalent neurodegenerative disorder worldwide, affecting approximately 56.9 million people in 2021 and projected to reach 152 million by 2050. Its defining pathological features, amyloid-beta plaques and neurofibrillary tangles, accumulate for up to two decades before cognitive symptoms emerge, placing the preclinical and mild cognitive impairment (MCI) stages at the centre of the early detection problem. Despite this, current diagnostic practice in routine clinical settings remains unreliable, with post-mortem studies placing the specificity of clinical AD diagnosis between 44.3 and 70.8% even in specialist memory clinics. Machine learning (ML) and deep learning (DL) applied to neuroimaging and electrophysiological data have emerged as candidate tools for closing this diagnostic gap, yet whether the accuracy figures reported in published studies translate into clinically useful performance on independent data remains unresolved. This study presents a structured comparative review of machine learning and deep learning methods reported across four publicly available Alzheimer’s disease datasets, namely the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Open Access Series of Imaging Studies (OASIS), the OpenNeuro ds004504 electroencephalography (EEG) dataset, and the Kaggle Alzheimer’s magnetic resonance imaging (MRI) dataset. Thirteen model families are examined through the published literature rather than through new experiments, and for each model and dataset combination, the best accuracy reported in the source study is recorded alongside the model’s mathematical formulation. All performance figures reported in this abstract and throughout the paper are taken from the published studies reviewed, not from new experiments conducted by the authors. Across the reviewed studies, deep learning architectures pre-trained on ImageNet and fine-tuned on neuroimaging data are reported to produce the highest accuracy on MRI classification tasks. Residual Network (ResNet)-101 is reported at 98.21 percent on ADNI and 97.45 percent on OASIS, while the IncepRes fusion architecture reaches 98.35% on OASIS by combining multi-scale feature extraction from InceptionV3 with residual connectivity from ResNet152V2. Traditional machine learning classifiers remain competitive on tabular clinical and biomarker data, with Extreme Gradient Boosting (XGBoost) reaching 91% on ADNI multiclass features. Logistic Regression achieves 82 to 85% on binary ADNI classification and is the only classifier in this review that provides explicit per-feature prediction contributions without post hoc tooling. Gaussian Naïve Bayes achieves 80 to 83% on the same task. On the OpenNeuro EEG dataset, K-nearest neighbours (KNN) with singular value decomposition (SVD) entropy features achieves 91% binary accuracy, with feature engineering quality determining performance more reliably than classifier architecture. Eight principal findings emerge from the cross-dataset analysis. Binary classification consistently outperforms multiclass by 10 to 30% across all datasets, reflecting the genuine biological ambiguity of the mild cognitive impairment category. Dataset size and augmentation predict reported accuracy more reliably than model architecture. Ensemble methods outperform individual classifiers by 5 to 8% in nearly every imaging study. Deeper architectures can overfit small clinical cohorts. EEG models trail MRI models by approximately 10 to 15% on comparable binary tasks. Cross-dataset generalisation has not been systematically evaluated in most studies, and the few that have tested it report accuracy drops of 5 to 10% or more when models encounter data from different scanners or cohorts. Eight recurring limitations constrain the clinical utility of these findings. Small sample sizes and limited demographic diversity, severe class imbalance inflating raw accuracy metrics, poor cross-dataset generalisation driven by scanner heterogeneity, limited deep learning interpretability, the dominance of binary over multiclass tasks, the absence of longitudinal modelling despite available datasets, inadequate standardisation of preprocessing and evaluation protocols, and the signal-to-noise ratio constraints specific to EEG recordings of elderly patients collectively define the gap between benchmark performance and clinical readiness. Future work must prioritise multi-centre training cohorts, multimodal fusion architectures, longitudinal progression modelling, and standardised interpretability evaluation as non-optional requirements for any system intended for clinical deployment. Full article

Artificial intelligence is increasingly used to support clinical decision making, yet concerns remain regarding algorithmic aversion, automation bias and the preservation of meaningful human oversight; while explainable AI aims to improve transparency, less attention has been devoted to the design of human–AI interaction protocols. This study investigates Frictional AI, an interaction paradigm that introduces cognitive friction to encourage critical engagement with AI recommendations. First, semi-structured interviews were conducted with a legal expert and a psychologist and analyzed through thematic analysis to identify legal, ethical, and cognitive requirements for AI-assisted decision support. Second, a user study involving 96 medical residents compared three interaction protocols: a conventional explainable AI-first design (XAI) and two friction-based protocols, namely a judicial protocol based on juxtaposed explanations (Judicial AI, JAI) and an adjunct protocol requiring an initial unsupported decision before AI exposure (AAI). Diagnostic accuracy and confidence, perceived usefulness, completion time, and reliance patterns were evaluated. The interviews highlighted the importance of human-centered explanations, contrastive reasoning, preservation of professional responsibility, and the role of user studies in evaluating human–AI interaction. The quantitative results showed that none of the AI-assisted conditions improved diagnostic accuracy relative to the no-support baseline. However, JAI achieved performance comparable to the baseline, outperforming XAI and AAI, and exhibited the lowest level of over-reliance. Overall findings suggest that the effectiveness of decision-support systems depends not only on model performance and explanation quality but also on interaction design. In conclusion, while preserving diagnostic performance, judicial protocols showed promise in mitigating automation bias and promoting active cognitive engagement in clinical decision support. Full article

Medical image classification in clinical settings is frequently constrained by High-Dimensional, Low-Sample Size (HDLSS) conditions, rendering conventional Support Vector Machines (SVM) geometrically susceptible to the data piling phenomenon. This study proposes a hybrid CNN-DWD framework to address this geometrical instability by integrating multi-architecture convolutional feature extraction with Distance-Weighted Discrimination (DWD). Pre-trained ResNet50 and DenseNet121 backbones act as frozen feature extractors, generating a highly descriptive 3072-dimensional fused representation. To resolve the computational bottleneck of deploying DWD directly on massive feature spaces, a supervised Multi-Layer Perceptron (MLP) bottleneck progressively compresses this space into a 32-dimensional latent manifold. Evaluated across breast ultrasonography, breast mammography, and chest X-ray datasets under varying training allocations, the proposed architecture drastically accelerates DWD training—achieving over a 130-fold speedup. The proposed CNN-MLP-DWD framework demonstrates highly competitive diagnostic performance, achieving 93.16% accuracy on the breast ultrasonography dataset and a macro-AUC of 99.69% on the chest X-ray benchmark, comparing favorably against the evaluated baseline methods. Full article

Recent language models exhibit significant logical reasoning abilities, yet the mechanisms supporting deductive inference remain poorly understood. This paper studies small transformer-based language models trained from scratch on multi-step deductive tasks, focusing on the distinction between horizontal reasoning, where intermediate steps are generated autoregressively, and vertical reasoning, where inference unfolds implicitly across layers before the first output token is produced. We analyze two synthetic tasks: logical consequence over chains of symbolic implications and root-to-leaf navigation in binary trees. Mechanistic interpretability reveals that Chain-of-Thought supervision enables models to learn rule-based inference rather than statistical shortcuts. In the horizontal setting, a shallow attention-only model develops interpretable circuits for rule completion, rule chaining, and final decision making, largely implemented through induction-head-like mechanisms. We further introduce a truncated pseudoinverse method to decode the information carried by queries, keys, and values. For vertical reasoning, Chain-of-Thought appears to act less as explicit step-by-step guidance and more as a form of curriculum learning, helping the model acquire increasingly complex reasoning patterns. Without Chain-of-Thought, models tend to memorize or exploit dataset biases. These results provide a low-level account of how transformers can implement deductive reasoning and suggest how Chain-of-Thought may serve different functions in horizontal and vertical reasoning. Full article

Predicting short-term progression in multiple sclerosis (MS) from longitudinal clinical data remains challenging because visits occur at irregular intervals, patient trajectories vary substantially, and simulated federated clients often exhibit non-IID data distributions. Privacy constraints further limit centralized data aggregation, reducing the suitability of conventional temporal prediction models. We propose a federated learning framework for MS progression prediction that combines local and global temporal modeling with client-level imbalance handling under simulated privacy-preserving federated constraints. The model uses a bidirectional LSTM to capture short-range temporal dynamics and a Transformer encoder to model longer-range dependencies in irregularly sampled sequences. A cross-attention mechanism fuses these representations. Each client applies conditional trajectory augmentation with ProxiWGANP and a proximity constraint. Training uses federated averaging with gradient clipping and Gaussian noise injection, while privacy loss is estimated through Rényi differential privacy accounting. We evaluated the framework on a single-center longitudinal cohort of 517 patients and 4184 visits. After excluding first visits without historical context and visits without determinable 6-month follow-up labels, we derived 3184 eligible prediction windows. We partitioned these windows into 10 simulated federated clients using Dirichlet sampling and evaluated performance with a strict within-patient temporal split. Across five independent training seeds, the proposed method achieved a mean AUROC of 0.930 with a standard deviation of 0.003 and a mean AUPRC of 0.513 with a standard deviation of 0.005. For the primary seed, patient-level bootstrap estimates were AUROC 0.930 (95% CI: 0.912–0.948), AUPRC 0.515 (95% CI: 0.490–0.540), and ECE 0.014. These results suggest that the proposed framework can support accurate short-term MS progression prediction in a simulated federated setting. However, because the evaluation used a single-center cohort and simulated federation, the findings provide methodological evidence and require external multi-center validation before clinical deployment. Full article

Replay-based continual learners are commonly evaluated at a fixed item count, but deployment hardware meters memory in bytes, and methods storing logits, feature vectors, or compressed codes occupy different byte footprints at an equal item count. The equal-item convention therefore does not answer the deployment question. We propose an equal-byte non-inferiority protocol that holds total stored bytes constant across competitors, selects the strongest of a pre-registered eight-method baseline pool at each budget, and reports a decision at a registered margin of ${\Delta }_{\mathrm{NI}}=0.05$ under Holm–Bonferroni correction. We demonstrate it on Split-CIFAR-100 and on an eight-task family-grouped sequence of fifty curated Abstraction and Reasoning Corpus puzzles, using a two-tier consolidation memory with a byte-budget knob, at ten seeds per cell. Under the pre-registered rule, the two-tier scheme fails non-inferiority at all four budgets on both benchmarks, the strongest per-buffer baseline exceeds the best two-tier cell at every budget, and the outcome is robust to task ordering on the abstract-reasoning benchmark. An interpretability analysis links the outcome to the compressed tier’s attention share and prototype similarity. The contribution is the protocol; the demonstration returns a clean negative decision on one specific candidate. Full article

Imbalanced datasets, where one class significantly outnumbers others, remain a persistent challenge in machine learning, often biasing predictions toward the majority class and degrading classifier performance. This paper provides a comprehensive, systematic review of data balancing methods, extending beyond foundational oversampling techniques such as the Synthetic Minority Oversampling Technique (SMOTE) and its variants (e.g., Borderline SMOTE, K-Means SMOTE, and Safe-Level SMOTE) to encompass advanced adaptive methods (MWMOTE, AMDO), deep generative models (generative adversarial networks, variational autoencoders, and diffusion models), undersampling techniques (NearMiss, Tomek Links), combination/hybrid methods (SMOTE-ENN, SMOTE-Tomek, and SMOTE+OCSVM), ensemble strategies (SMOTEBoost, RUSBoost, Balanced Random Forest, and One-Sided Selection), and specialized approaches for multi-label and clustered data. Beyond descriptive categorization, this review critically examines each method’s underlying assumptions, operational mechanisms, and suitability for diverse data characteristics, including high dimensionality, mixed feature types, class overlap, and noise. Key findings demonstrate that no single method universally outperforms others; optimal selection depends critically on dataset characteristics, classifier choice, and evaluation metrics. The paper concludes by identifying emerging research directions, including self-supervised learning for imbalance, diffusion-based generative oversampling, distribution-preserving resampling, knowledge distillation for imbalanced deployment, and the adaptation of foundation models to skewed distributions, offering practical guidelines for practitioners and a roadmap for future methodological development. Full article

Digital twins (DTs) have become a central paradigm for modeling cyber–physical systems and digital infrastructures, yet the term is applied to very different representations—from physical assets to operational processes and service environments. This ambiguity obscures how the various DT interpretations relate to one another and at which level knowledge can be represented and extracted. This paper develops a conceptual and mathematical framework that treats asset-centric, process-centric, and service-centric DTs as successive levels of system abstraction. DTs are modeled as mappings between real-world entities and their digital representations, and the three paradigms are connected through explicit cross-layer dependencies, with service-centric twins shown to form a distinct level that cannot be reduced to asset and process descriptions alone; the framework is then extended to the ecosystem level as a digital service ecosystem twin. Because each level fixes the entities, features, and relations available to data-driven methods, the framework also specifies where machine-learning and knowledge-extraction tasks operate within layered DT architectures. The approach is illustrated and validated for structural and cross-layer consistency through a smart-city electricity ecosystem, providing a unified basis for interpreting the evolution of DTs toward service-oriented digital ecosystems. Full article

Copper is a strategically important commodity whose price dynamics are increasingly affected by structural changes, geopolitical shocks, and the global energy transition. These conditions create substantial challenges for forecasting models and provide a useful setting for evaluating the practical value of machine learning predictions. This study compares statistical and artificial intelligence-based forecasting models for copper price prediction under different market regimes and structural break conditions. Model performance is assessed using a multi-dimensional evaluation framework that combines statistical accuracy (MAPE), dynamic pattern reproduction (Taylor diagrams and time-lagged cross-correlation analysis), and the economic performance of forecast-driven trading strategies. The results reveal a consistent error–profit paradox: models with the highest statistical forecasting accuracy do not necessarily generate the best trading outcomes. In several cases, models with larger prediction errors achieve superior economic performance because they capture directional market dynamics more effectively. The analyses further show that structural breaks substantially alter model rankings and predictive usefulness, highlighting the importance of regime-aware evaluation. These findings suggest that forecast accuracy alone provides an incomplete assessment of model quality in financial and commodity forecasting applications. The study contributes to machine learning evaluation research by proposing an integrated framework that jointly considers predictive accuracy, temporal dynamics, model robustness, and economic utility, thereby offering a more comprehensive approach to assessing forecasting systems in real-world decision-making environments. Full article

While handwritten signatures are a staple of biometric authentication, conventional verification models typically rely on Euclidean space assumptions, restricting the capture of complex, intrinsic signature structures. To address this, offline signature verification has increasingly modeled signatures as points on the Symmetric Positive Definite (SPD) manifold. Nevertheless, selecting an appropriate metric on this manifold for a given problem remains a significant challenge, typically relying on heuristic trial-and-error processes. To solve this, our primary contribution is a novel, end-to-end Riemannian framework featuring the Collective Background Knowledge (CBK) mechanism. CBK establishes synthetic writers as Riemannian centers, utilizing a learnable $\alpha \beta$-Log-Determinant divergence to adaptively discover the optimal local geometry from data. Instead of computing the direct distance between two SPD signature representations, we evaluate them relationally by measuring how each signature diverges from the shared CBK reference centers. These individual deviations form unique relational profiles for each signature, which are then compared using the dichotomy transform to create a dissimilarity vector. By jointly optimizing the CBK parameters under an SPD metric-learning approach, our model separates effectively similar and dissimilar pairs of signatures. Evaluated across five datasets under challenging blind intra- and cross-lingual conditions, our geometry-aware framework demonstrates robust generalization and competitive performance. Full article

Frontier large language models achieve broad linguistic competence but degrade on specialist domains underrepresented in pre-training corpora. Domain-adaptive post-training (DAPT) on curated professional text partially closes this gap, yet the dominant approach flattens structured discourse into isolated document units, discarding the collaborative reasoning embedded in multi-party exchanges. This paper investigates whether preserving the full recursive structure of user forum threads during post-training is a more effective first step toward knowledge extraction than flattened question-answer pairs. Four open-source decoder-only models (TinyLlama 1.1B, Phi-2 2.7B, LLaMA-2-7B 6.8B, LLaMA-2-13B 13B) are post-trained using parameter-efficient LoRA adaptation on 4970 threads from AgTalk, an agricultural producer forum, under three conditions: flattened Q → A pairs, full recursive threads preserving reply order, and shuffled recursive threads with randomly permuted intermediate replies. Five hypotheses are tested through paired Wilcoxon signed-rank comparisons across 29 thread positions. DAPT significantly reduces perplexity relative to pretrained baselines across all architectures (H0 supported). Recursive training reduces perplexity relative to flattened training, an advantage clearest for the two LLaMA-2 models under matched-context evaluation (Wilcoxon win rates near 72%) and present but obscured by outlier variance at the 1.1B and 2.7B scales (H₁ supported). However, ordered recursive training provides only a marginal advantage over shuffled (H₂ inconclusive), attention analysis reveals identical U-shaped endpoint-weighted profiles regardless of training condition (H₃: architectural not learned), and perplexity shows no systematic decrease with accumulating thread depth (H₄ not supported). These results are attributed to Rotary Position Embedding constraints in decoder-only architectures that systematically underweight middle-thread content. Encoder–decoder architectures with bidirectional cross-attention are identified as a promising next step for exploiting the full collaborative structure of forum discourse. Full article

Detecting localized morphological anomalies in three-dimensional point clouds is difficult because geometric deviations are entangled with rigid pose variation, residual registration error, sampling noise, and normal inter-subject variability. This challenge is particularly relevant in translational neuroimaging, where abnormal shape changes may be subtle and abnormal annotations are scarce. We propose an unsupervised framework that formulates 3D anomaly detection as a two-stage factorization problem, termed AdvFlow3D-AD. First, Fast Global Registration, followed by multi-scale Iterative Closest Point refinement, establishes a common geometric reference frame and reduces rigid-body nuisance variation. Second, an adversarially regularized normalizing flow models the residual distribution of aligned normal coordinates, enabling localized anomaly scores based on distance from the learned normal latent support. Percentile calibration on normal data then defines interpretable point-level and object-level operating points without requiring abnormal samples during training. We evaluate AdvFlow3D-AD on the Real3D-AD and Anomaly ShapeNet3D datasets, achieving a point-level area under the receiver operating characteristic curve (AUROC) of 0.747 on Real3D-AD and an object-level AUROC of 0.816 on Anomaly ShapeNet3D. We further present an exploratory neurodevelopmental brain-shape case study involving pediatric perinatal-asphyxia cases. The resulting anomaly maps showed qualitative spatial correspondence with anatomically plausible hippocampal and cerebellar regions under neuroradiological review. These results suggest that separating geometric nuisance variation from residual morphology can support interpretable anomaly localization when abnormal labels are limited. Full article

This study investigates automated fake news detection as a reliability-oriented text classification problem in dynamic digital information environments. We propose H-FuseNet, a hybrid multi-representation fusion framework that combines pretrained transformer representations with deception-oriented handcrafted linguistic, stylistic, and semantic features. Using the WELFake dataset, we benchmark 15 baseline models, including classical classifiers, ensemble methods, recurrent and convolutional networks, and transformer fine-tuning models, under stratified 10-fold cross-validation with nested hyperparameter optimization. To examine generalization beyond a single benchmark, we train exclusively on WELFake and evaluate cross-dataset performance on three held-out external datasets: FakeNewsNet, CoAID, and LLM-generated misinformation. H-FuseNet integrates transformer document embeddings with a lightweight feature-processing MLP, optional contextual feature streams when metadata are available, and auxiliary supervision through pseudo-labeled headline body stance and clickbait signals. The proposed model achieves 98.9% mean accuracy and 0.998 ROC–AUC, while maintaining strong calibration, with a Brier score of 0.012 and Expected Calibration Error of 0.009, and low variance across folds. Cross-dataset evaluation yields accuracies of 87.34% on FakeNewsNet, 83.56% on CoAID, and 91.22% on LLM-generated misinformation, demonstrating robust generalization under distribution shift. Ablation analyses show that handcrafted features, auxiliary tasks, and learned fusion each contribute to performance, while Wilcoxon and McNemar tests indicate statistically significant differences against selected strong baselines. Error analysis shows that remaining failures mainly occur in professionally written misinformation that imitates neutral journalistic style. Overall, the results suggest that calibrated multi-representation fusion can improve the reliability of automated fake news detection systems. Full article

The YOLO (You Only Look Once) object detection models have undergone rapid evolution, with each version introducing architectural enhancements aiming to improve speed, accuracy, and deployment. Simultaneously, single-board computers (SBCs) have advanced to support increasingly complex AI models in edge environments. This study presents a comprehensive benchmarking of YOLO versions 8 through 12 across a range of SBCs, including Raspberry Pi4/5, NVIDIA Jetson Nano, Jetson Orin, and LattePanda, under different power modes. Key performance metrics, including inference speed (FPS), detection accuracy (mAP), RAM usage, and computational complexity (FLOPs), are evaluated. These findings offer practical insights for developers and researchers to select optimal YOLO variants and SBC configurations for real-time edge deployment. Full article

We propose a modular, retrieval-augmented pipeline for computational argumentation that integrates two complementary components: ArgStance, a multi-task model for argument and stance reasoning, and TargetMatch, a contrastive retrieval model that treats target identification as a first-class retrieval task. We further formulate stance detection as a framing-aware problem, recognizing that the polarity of a stance toward a target depends on how the proposition is framed. To support broad generalization, we construct a large dataset spanning Kialo discussions, Wikipedia, and curated news articles, and introduce a cross-source injection strategy that mitigates domain and style biases. Our compact models achieve F1 scores of 0.94 for argument detection (ModernBERT-base) and 0.84 for same-side stance detection (ModernBERT-large), while TargetMatch attains a top-10 retrieval accuracy of 0.75. Under controlled zero-shot comparisons with large language models, our models remain competitive while offering advantages in reproducibility, deployment cost, and controllable intermediate predictions. Full article

Background: Oral cavity tumors are often diagnosed at advanced stages due to non-specific early symptoms and limitations in imaging sensitivity. This study proposes a novel methodology combining Radiomics and Model Checking to improve diagnosis and prognosis. Methods: A retrospective dataset of 18 patients (12 with oral squamous cell carcinoma and 6 healthy controls) who underwent contrast-enhanced MRI was analyzed. Radiomic features were extracted and selected, then encoded into formal models. A two-tier Model Checking approach was applied: (i) classification of healthy vs. pathological patients and (ii) prediction of treatment response. Results: The proposed method achieved a diagnostic accuracy of 93% and a prognostic accuracy of 75%. The approach demonstrated robustness even with a limited dataset, outperforming traditional data-driven methods in small-sample settings. Conclusions: The integration of Radiomics with Model Checking provides an explainable and effective tool for early detection and prognosis of oral cavity tumors. This approach may support clinicians as a decision-making aid, particularly in data-scarce scenarios. Full article

Large language models can emit fluent text while leaving intermediate semantic structure implicit. We study whether explicit event-role and logical-primitive side-state can improve a pretrained decoder without damaging its language behavior. We introduce PAT-ER, a decoder architecture with a normal token stream, an event-role register stream, and a primitive register stream. The primitive stream is motivated by the view that logical primitives answer characteristic semantic questions, such as what licenses a conclusion, what conflicts with it, or why evidence is insufficient. Across eight seeds on the same Qwen3-0.6B backbone, replacing token-pooled auxiliary heads with typed PAT-ER registers improves primitive macro-F1 by 0.209 (95% CI [0.182, 0.237]) and role-to-primitive macro-F1 by 0.091 (95% CI [0.074, 0.110]) with no language-model loss cost. A generic-register control shows that this is not merely the effect of adding latent registers: typed PAT-ER improves over generic registers by 0.116 primitive macro-F1 and 0.110 role-to-primitive macro-F1, with both confidence intervals excluding zero. A warm-started model then recovers pretrained language quality (LM loss 1.344 versus 2.555 for the frozen-backbone register model) while retaining most side-state behavior. Finally, a decoupled interface mode produces robust schema-grounded function calls on 242 held-out prompts (Hermes parse 0.952, exact arguments 0.981, JSON validity 1.000, IDK F1 1.000) while base-mode side-state metrics remain byte-identical to the warm-start baseline. The model is not a theorem prover and does not achieve perfect unseen tool-name copying; the contribution is a measured architecture signal and a usable, guarded interface. Full article

This study presents a comprehensive, evidence-based methodological framework for estimating extreme significant wave heights associated with several return periods using meteorological records (wind speed and air temperature) in combination with machine learning algorithms. The proposed framework integrates a sequence of preliminary diagnostic analyses prior to extreme value modelling, including homogeneity tests to identify potential change points and trend analyses to assess the presence of stationary or non-stationary behaviour, thereby enabling the selection of an appropriate frequency analysis approach for computing annual maximum significant wave heights across several return periods. The methodology is applied to annual maximum significant wave height records obtained from offshore buoy stations in the Antilles region of the Caribbean Sea. Results from the machine learning presets demonstrate a strong relationship between extreme significant wave heights and the corresponding wind speed and temperature records, yielding coefficients of determination (R²) of 0.75 and 0.88 for the validation and testing stages, respectively. When frequency analysis is conducted using both the traditional approach and the proposed machine learning–based methodology, high agreement is observed for return periods between 2 and 20 years, with R² values ranging from 0.98 to 0.89. For longer return periods, the agreement decreases, which is consistent with the limited length of the available wave height records and the supervised nature of the machine learning algorithms. In this sense, the proposed framework provides a robust alternative for estimating extreme significant wave heights in regions where direct wave observations are unavailable but meteorological records are accessible. Full article

Autonomous driving systems rely on a sophisticated pipeline of artificial intelligence models to perceive, predict, and plan in dynamic environments. This review presents a systematic analysis of the machine learning and deep learning models underpinning vehicle autonomy, spanning classical convolutional neural networks (CNNs) for object detection and semantic segmentation to recurrent and Transformer-based architectures for trajectory prediction and motion planning. It also provides a critical examination of the autonomous vehicle sensor stack, including cameras, LiDAR, radar, ultrasonics, and GNSS/IMU as data acquisition systems, highlighting modality-specific AI challenges such as monocular depth estimation, 3D point cloud processing, and radar Doppler interpretation. The evolution of perception and decision-making pipelines is reviewed, contrasting modular architectures with end-to-end learning paradigms that directly map raw sensor data to control commands, and discussing their trade-offs in interpretability, safety assurance, and robustness to rare edge cases. We further survey specialized hardware accelerators and heterogeneous automotive SoCs designed to meet stringent real-time and power constraints. Industrial strategies are compared, including multi-modal sensor fusion and vision-centric approaches based on large-scale imitation learning. Finally, we identify open challenges related to robustness under adverse conditions, domain shift, causal ambiguity, and the need for interpretable and certifiable AI in safety-critical autonomous driving systems. Full article

Prostate-specific antigen (PSA) is a well-established marker for prostate cancer screening, but current ≥4 ng/mL cutoff suffers from low specificity. This study aims to demonstrate the use of a long short-term memory (LSTM) network for accurate prediction of prostate cancer risk and next sequential PSA value. Hong Kong-wide PSA test data over a 25-year period, including PSA values, time difference between PSA tests, and PSA velocity change, were retrieved for model training with variable PSA cutoffs and sequence length. A total of 1,158,915 PSA tests from 499,342 patients (including 18,629 patients with prostate cancer) were included. Models predicting the next PSA level performed well (accuracy 0.724–0.910, AUROC 0.751–0.892). For a ≥4 ng/mL cutoff, the best model was at sequence length of 4 (AUROC 0.864). Temporal prediction of PSA performance was lower (accuracy 0.739–0.811, AUROC 0.740–0.849). Prostate cancer prediction performed excellent (AUROC 0.888–0.973), the sensitivity (0.734), and specificity (0.962) were high even at the shortest sequence length and a ≥4 ng/mL cutoff (4). Utilizing PSA levels only without additional markers or clinical data, LSTM-based models accurately predicted the next PSA level with modest temporal predictions while significantly improving the specificity of prostate cancer risk prediction, demonstrating their clinical utility. Full article

The marine capture fisheries sector of the Philippines employs approximately 2.3 million Filipinos, yet recent declines (including a 15.3% drop in Q1 2026 production relative to Q1 2025) underscore the need for forecasting systems resolved at the regional and sectoral level. Existing Philippine approaches rely on univariate classical time-series methods and seldom integrate multivariate oceanographic predictors. This study addresses three questions: (RQ1) How do nine candidate machine learning algorithms compare in forecasting regional fish production from environmental predictors? (RQ2) Which environmental predictors most strongly drive model output, as quantified by explainable AI (XAI) SHAP-based feature attribution? (RQ3) To what extent do model performance and predictor importance vary across regions? Across 32 region–sector panels spanning 2002–2025, kernel and neural network models were selected as the best-performing architecture in 26 of 32 panels (81.3%), achieving a mean composite score 12.7% higher than tree-based ensembles, a gap attributable to extrapolation along trending physical predictors. Feature attribution identified the partial pressure of CO₂ as the leading driver in both sectors, exceeding the second-ranked variable by factors of 2.5 (commercial) and 3.4 (marine municipal). Regional heterogeneity in retained predictors, winning algorithms, and SHAP attribution rankings supports region-specific forecasting as a necessary design choice. Mean absolute percentage error of 22–25% and directional accuracy of 0.62–0.66 indicate operational utility for early-warning applications, establishing a basis for evidence-driven priority-setting in Philippine fisheries governance. Full article

The creation of collaborative knowledge bases for decision support systems (DSS) mitigates the subjectivity of individual experts and enhances overall system efficacy. However, traditional knowledge engineering approaches are highly labor-intensive when eliciting and integrating expert knowledge, as they require extensive, expert-level interaction between knowledge engineers and domain specialists. Modern large language models (LLMs) and retrieval-augmented generation (RAG) technologies present novel opportunities for overcoming these limitations. This study presents a pilot investigation to assess the potential of LLM-based knowledge engineering for developing collaborative knowledge bases within knowledge-based DSS, thereby assisting decision-making in complex operational scenarios involving technical systems. The article proposes an LLM-based approach for creating collaborative knowledge bases, including extraction, consolidation of expert knowledge and evaluation of their operability. To implement and evaluate the proposed approach, specialized prompts were engineered, and pilot experiments were conducted to generate consolidated knowledge cases through expert-LLM interactions. The resulting knowledge cases were subsequently applied in an experimental decision-making inference procedure for fault diagnosis in gas-fired heating boilers. During this inference process, an LLM agent, guided by tailored prompts and a RAG-enabled knowledge base, interactively queries the user to identify the specific issue and subsequently proposes a contextually appropriate solution. Throughout this study, the LLMs demonstrated capabilities in dialogue management, expert knowledge elicitation, and knowledge consolidation, successfully facilitating the creation of a collaborative knowledge base grounded in the “Event-Cause-Symptoms-Action” model. The findings highlight the viability of future research in LLM-based knowledge engineering and support the further advancement of the “LLM-as-knowledge-engineer” paradigm. Full article

Interpretable machine learning for ordinal quality grading faces a fundamental tension between model transparency and predictive performance, particularly under small-data conditions where end-to-end deep learning is unreliable and domain knowledge must compensate for limited training samples. We present a dual-target feature engineering framework for interpretable ordinal grading validated on pane Carasau, a traditional flatbread whose extreme surface variability makes it a challenging small-data benchmark for machine learning under realistic acquisition constraints. The pipeline extracts 116 handcrafted visual descriptors organised into four families—colour, texture, spatial, and hotspot—and grades the quality along two independent axes: global toasting intensity and spatial uniformity, complemented by a continuous Toasting Index for process monitoring, on a dataset of 1512 images spanning four acquisition campaigns and three product types. On the primary within-batch evaluation set Campaign 01, $N=1090$), XGBoost achieves F1 macro $=0.906$ for toast classification and $R^2=0.886$ for continuous regression, substantially outperforming two fine-tuned CNN baselines on the same evaluation set (MobileNetV2: F1 $=0.523$; EfficientNet-B0: F1 $=0.518$). Feature importance analysis reveals that colour descriptors dominate toasting prediction ($87.5\%$), whilst spatial and texture features are essential for uniformity assessment ($47.4\%$ combined), providing physically grounded explanations directly traceable to the underlying thermal process. Cross-batch generalisation on held-out campaigns is moderate for the same product (XGB F1 = 0.718, $\kappa$ = 0.703); cross-product transfer to geometrically distinct variants requires product-specific adaptation. The framework requires no GPU, runs on standard CPU hardware at 4 s per image, and provides complete decision transparency, supporting deployment without specialised hardware. Full article

Graph machine learning reports steady progress across node, graph, and link prediction, across temporal and hypergraph frontiers, and across the emerging class of graph foundation models. This review asks a prior question: when a method is reported to outperform the alternatives, how far does the evidence support the claim? We organize the answer around four axes of evaluation rigor: statistical rigor (seeds, dispersion, formal significance testing), baseline fairness (budget-parity tuning of trivial and structure-agnostic baselines), data integrity (leakage, duplication, negative sampling, contamination), and claim integrity (whether gains survive fair tuning and discriminative benchmarks). Drawing on a criterion-based corpus of 150 studies, of which 51 were read in full depth, we find a consistent picture. Only two of 25 methodologically central backbone studies apply a formal between-method significance test, and reported gains repeatedly shrink or disappear once a trivial baseline is tuned to parity, a leaked split is repaired, or a pretrained model is evaluated on unseen data. We argue that these failures share one cause: the saturation of benchmarks that can no longer discriminate between methods. The principal output is a minimum reporting standard, a concrete four-axis checklist that authors and reviewers can apply at negligible cost. Full article

Gaussian process-based Bayesian optimization (BO) is a sample-efficient sequential strategy for optimizing expensive black-box functions. The Gaussian process provides a probabilistic approximation of the unknown function, while an acquisition function balances exploration and exploitation to select the next evaluation point. Despite significant research efforts, no master acquisition function has been identified. This paper proposes a novel adaptive acquisition function that dynamically adjusts the exploration–exploitation trade-off based on the evolution of the optimization process, rather than using fixed or random scheduling. While implemented here within a GP-based BO framework, the core switching mechanism is surrogate-agnostic: the exploitative component requires only a surrogate point prediction, and the explorative component is entirely model-free. Unlike traditional approaches, where mechanisms like UCB/LCB lean toward exploration over iterations, or fixed strategies that switch from exploratory (EI) to exploitative (PI) behavior at predetermined points, the proposed method makes purely exploitative decisions using only the GP’s prediction. However, it discards these decisions when they have low potential for significant improvement, instead focusing on uncertainty reduction. Notably, this approach uses inverse distance weighting for uncertainty quantification rather than the GP’s predictive uncertainty, avoiding bias from the GP’s predictions. Testing on benchmark functions demonstrates that the proposed acquisition function is almost always Pareto optimal, offering the most balanced trade-off between convergence to the global optimum and exploration capability compared to state-of-the-art alternatives. Full article

Selecting machine learning algorithms for clinical deployment demands comprehensive evaluation beyond conventional performance metrics. While automated frameworks simplify model generation, identifying algorithms suitable for real-world medical applications requires systematic assessment of learning dynamics, generalization stability, and cross-subset reliability. This study addresses prediabetes prediction through a multi-stage evaluation comparing automated machine learning frameworks, neural networks, gradient boosting implementations (XGBoost, CatBoost, LightGBM), and specialized imbalance-handling techniques. A questionnaire-based dataset with a substantial class imbalance was analyzed through progressive evaluation stages: aggregate performance metrics, learning curve analysis, minority class detection capability, and cross-subset generalization stability. Linear Discriminant Analysis achieved maximum validation metrics in automated screening but exhibited flat learning curves, indicating an exhausted learning capacity. XGBoost demonstrated optimal convergence dynamics with the highest validation performance (0.749 AUC), yet suffered substantial validation-to-test degradation (5.9 percentage points). CatBoost, despite inferior validation performance (0.696 accuracy), exhibited exceptional cross-subset stability with minimal performance decline (0.2 percentage points) while achieving a comparable test accuracy (0.694). CatBoost was selected for deployment based on its superior generalization stability, demonstrating that a multi-dimensional evaluation spanning aggregate metrics, learning dynamics, and cross-subset stability is essential for identifying clinically deployable models, as validation performance alone provides insufficient evidence for real-world applicability. Full article

Information technology service management (ITSM) systems generate large volumes of unstructured incident descriptions. They frequently include multilingual content, code-switching, informal language, and domain-specific terminology. These characteristics make automated text processing substantially more complicated and limit the applicability of conventional machine translation solutions, particularly in environments subject to strict data privacy and air-gap constraints. This paper presents a system-level reproducibility study of a deterministic two-stage cascaded neural machine translation (NMT) pipeline for normalizing multilingual IT incident text in resource-constrained, air-gapped environments. The study evaluates a sequential RU→EN and LT→EN translation strategy specifically selected to bypass unreliable language identification, enabling stable processing of code-switched incident descriptions. A system-level processing pipeline, which includes text normalization, segmentation, deduplication, adaptive batching, and language-aware data flow optimization, is analyzed to assess its impact on reducing redundant inference operations. The methodology is evaluated on a real-world ITSM dataset comprising 84,285 incident records. An incremental experimental design is used to isolate the specific contributions of computational and data-flow optimizations. Translation quality is assessed using BLEU and COMET metrics against expert reference translations produced via a primary translation and subsequent cross-verification by a second domain expert to ensure linguistic and technical consistency. The results indicate that a cascaded NMT architecture combined with systematic data-flow optimization provides a reproducible and privacy-preserving framework for multilingual IT incident text normalization, effectively supporting downstream analytical tasks in constrained operational ITSM environments. Full article

Anomaly detection systems deployed in critical applications require not only high accuracy but also reliable uncertainty quantification and coverage guarantees. This paper is an empirical study—rather than a contribution of new conformal-prediction machinery—of class-conditional (Mondrian) conformal prediction for anomaly detection under extreme class imbalance, characterizing where standard conformal prediction fails and how class-conditional calibration restores valid coverage. Class-conditional conformal prediction constructs prediction sets that, under exchangeability, contain the true label with user-specified confidence (e.g., 90%), enabling systems to abstain on uncertain predictions. Unlike standard conformal prediction that fails catastrophically under extreme imbalance—achieving only 52.94% anomaly coverage at a 1:345 imbalance ratio—class-conditional calibration maintains 90.59% anomaly coverage by computing quantiles separately for each class. We apply the standard softmax-based nonconformity score $s=1-f_y\left(x\right)$ within each class, ensuring valid coverage for both normal and anomalous instances with coverage gaps ranging from 0.50% to 5.18% depending on dataset characteristics. Extensive experiments on three real-world datasets (Microsoft Azure KPI, Yahoo, NAB) demonstrate that the method achieves empirical coverage within 0.06–0.33% of theoretical targets at confidence levels $\alpha \le 0.05$; on the most imbalanced benchmark (Microsoft Azure KPI at a 1:345 ratio and $\alpha =0.10$), this corresponds to a 37.65 percentage point improvement in anomaly coverage over standard conformal prediction. We restate finite-sample coverage bounds and exchangeability conditions in the binary anomaly detection setting and validate them empirically through Monte Carlo simulation. Multi-model evaluation across XGBoost, Random Forest, and Neural Networks demonstrates the model-agnostic property of the framework, while also identifying conditions (poor base-classifier discrimination, small minority calibration sets) under which coverage may be marginally violated. Comparison with alternative uncertainty quantification methods (isotonic probability calibration, Monte Carlo dropout) shows that only conformal prediction provides formal guarantees while maintaining 90.59% anomaly coverage versus 76.47% and 84.71% for alternatives. The abstention mechanism identifies 34–66% of predictions as uncertain at high confidence levels (99%), enabling safety-critical systems to defer difficult cases to human experts while preserving baseline discrimination (ROC-AUC unchanged). Full article