Mach. Learn. Knowl. Extr., Volume 7, Issue 1 (March 2025) – 26 articles

Equatorial plasma bubbles (EPBs) are regions of depleted electron density that form in the Earth’s ionosphere due to Rayleigh–Taylor instability. These bubbles can cause signal scintillation, leading to signal loss and errors in position calculations. EPBs can be detected in images captured by All-Sky Imager (ASI) systems. This study proposes a low-cost automatic detection method for EPBs in ASI data that can be used for both real-time detection and classification purposes. This method utilizes Two-Dimensional Principal Component Analysis (2DPCA) with Recursive Feature Elimination (RFE), in conjunction with a Random Forest machine learning model, to create an Explainable Artificial Intelligence (XAI) model capable of extracting image features to automatically detect EPBs with the lowest possible dimensionality. This led to having a small-sized and extremely fast-trained model that could be used to identify EPBs within the captured ASI images. A set of 2458 images, classified into two categories—Event and Empty—were used to build the database. This database was randomly split into two subsets: a training dataset (80%) and a testing dataset (20%). The produced XAI model demonstrated slightly higher detection accuracy compared to the standard 2DPCA model while being significantly smaller in size. Furthermore, the proposed model’s performance has been evaluated and compared with other deep learning baseline models (ResNet18, Inception-V3, VGG16, and VGG19) in the same environment. Full article

Non-negative Matrix Factorization (NMF) has gained popularity due to its effectiveness in clustering and feature selection tasks. It is particularly valuable for managing high-dimensional data by reducing dimensionality and providing meaningful semantic representations. However, traditional NMF methods may encounter challenges when dealing with noisy data, outliers, or when the underlying manifold structure of the data is overlooked. This paper introduces an innovative approach called SGRiT, which employs Stiefel manifold optimization to enhance the extraction of latent features. These learned features have been shown to be highly informative for clustering tasks. The method leverages a spectral decomposition criterion to obtain a low-dimensional embedding that captures the intrinsic geometric structure of the data. Additionally, this paper presents a solution for addressing the Stiefel manifold problem and utilizes a Riemannian-based trust region algorithm to optimize the loss function. The outcome of this optimization process is a new representation of the data in a transformed space, which can subsequently serve as input for the NMF algorithm. Furthermore, this paper incorporates a novel subspace graph regularization term that considers high-order geometric information and introduces a sparsity term for the factor matrices. These enhancements significantly improve the discrimination capabilities of the learning process. This paper conducts an impartial analysis of several essential NMF algorithms. To demonstrate that the proposed approach consistently outperforms other benchmark algorithms, four clustering evaluation indices are employed. Full article

The electrical storm optimization (ESO) algorithm, inspired by the dynamic nature of electrical storms, is a novel population-based metaheuristic that employs three dynamically adjusted parameters: field resistance, field intensity, and field conductivity. Field resistance assesses the spread of solutions within the search space, reflecting strategy diversity. The field intensity balances the exploration of new territories and the exploitation of promising areas. The field conductivity adjusts the adaptability of the search process, enhancing the algorithm’s ability to escape local optima and converge on global solutions. These adjustments enable the ESO to adapt in real-time to various optimization scenarios, steering the search toward potential optima. ESO’s performance was rigorously tested against 60 benchmark problems from the IEEE CEC SOBC 2022 suite and 20 well-known metaheuristics. The results demonstrate the superior performance of ESOs, particularly in tasks requiring a nuanced balance between exploration and exploitation. Its efficacy is further validated through successful applications in four engineering domains, highlighting its precision, stability, flexibility, and efficiency. Additionally, the algorithm’s computational costs were evaluated in terms of the number of function evaluations and computational overhead, reinforcing its status as a standout choice in the metaheuristic field. Full article

This study investigates the integration of diverse data modalities within deep learning ensembles for Android malware classification. Android applications can be represented as binary images and function call graphs, each offering complementary perspectives on the executable. We synthesise these modalities by combining predictions from convolutional and graph neural networks with a multilayer perceptron. Empirical results demonstrate that multimodal models outperform their unimodal counterparts while remaining highly efficient. For instance, integrating a plain CNN with 83.1% accuracy and a GCN with 80.6% accuracy boosts overall accuracy to 88.3%. DenseNet-GIN achieves 90.6% accuracy, with no further improvement obtained by expanding this ensemble to four models. Based on our findings, we advocate for the flexible development of modalities to capture distinct aspects of applications and for the design of algorithms that effectively integrate this information. Full article

We introduce a weakly supervised segmentation approach that leverages class activation maps and the Segment Anything Model to generate high-quality masks using only classification data. A pre-trained classifier produces class activation maps that, once thresholded, yield bounding boxes encapsulating the regions of interest. These boxes prompt the SAM to generate detailed segmentation masks, which are then refined by selecting the best overlap with automatically generated masks from the foundational model using the intersection over union metric. In a polyp segmentation case study, our approach outperforms existing zero-shot and weakly supervised methods, achieving a mean intersection over union of 0.63. This method offers an efficient and general solution for image segmentation tasks where segmentation data are scarce. Full article

The growing sophistication of cyber threats necessitates robust and interpretable intrusion detection systems (IDS) to safeguard network security. While machine learning models such as Decision Tree (DT), Random Forest (RF), k-Nearest Neighbors (K-NN), and XGBoost demonstrate high effectiveness in detecting malicious activities, their interpretability decreases as their complexity and accuracy increase, posing challenges for critical cybersecurity applications. Local Interpretable Model-agnostic Explanations (LIME) is widely used to address this limitation; however, its reliance on normal distribution for perturbations often fails to capture the non-linear and imbalanced characteristics of datasets like CIC-IDS-2018. To address these challenges, we propose a modified LIME perturbation strategy using Weibull, Gamma, Beta, and Pareto distributions to better capture the characteristics of network traffic data. Our methodology improves the stability of different ML models trained on CIC-IDS datasets, enabling more meaningful and reliable explanations of model predictions. The proposed modifications allow for an increase in explanation fidelity by up to 78% compared to the default Gaussian approach. Pareto-based perturbations provide the best results. Among all distributions tested, Pareto consistently yielded the highest explanation fidelity and stability, particularly for K-NN ($R^2$ = 0.9971, S = 0.9907) and DT ($R^2$ = 0.9267, S = 0.9797). This indicates that heavy-tailed distributions fit well with real-world network traffic patterns, reducing the variance in attribute importance explanations and making them more robust. Full article

This research paper presents novel approaches for detecting credit card risk through the utilization of Long Short-Term Memory (LSTM) networks and XGBoost algorithms. Facing the challenge of securing credit card transactions, this study explores the potential of LSTM networks for their ability to understand sequential dependencies in transaction data. This research sheds light on which model is more effective in addressing the challenges posed by imbalanced datasets in credit risk assessment. The methodology utilized for imbalanced datasets includes the use of the Synthetic Minority Oversampling Technique (SMOTE) to address any imbalance in class distribution. This paper conducts an extensive literature review, comparing various machine learning methods, and proposes an innovative framework that compares LSTM with XGBoost to improve fraud detection accuracy. LSTM, a recurrent neural network renowned for its ability to capture temporal dependencies within sequences of transactions, is compared with XGBoost, a formidable ensemble learning algorithm that enhances feature-based classification. By meticulously carrying out preprocessing tasks, constructing competent training models, and implementing ensemble techniques, our proposed framework demonstrates unwavering performance in accurately identifying fraudulent transactions. The comparison of LSTM and XGBoost shows that LSTM is more effective for our imbalanced dataset. Compared with XGBOOST’s 97% accuracy, LSTM’s accuracy is 99%. The final result emphasizes how crucial it is to select the optimal algorithm based on particular criteria within financial concerns, which will ultimately result in more reliable and knowledgeable credit score decisions. Full article

Explainability is essential for AI models, especially in clinical settings where understanding the model’s decisions is crucial. Despite their impressive performance, black-box AI models are unsuitable for clinical use if their operations cannot be explained to clinicians. While deep neural networks (DNNs) represent the forefront of model performance, their explanations are often not easily interpreted by humans. On the other hand, hand-crafted features extracted to represent different aspects of the input data and traditional machine learning models are generally more understandable. However, they often lack the effectiveness of advanced models due to human limitations in feature design. To address this, we propose ExShall-CNN, a novel explainable shallow convolutional neural network for medical image processing. This model improves upon hand-crafted features to maintain human interpretability, ensuring that its decisions are transparent and understandable. We introduce the explainable shallow convolutional neural network (ExShall-CNN), which combines the interpretability of hand-crafted features with the performance of advanced deep convolutional networks like U-Net for medical image segmentation. Built on recent advancements in machine learning, ExShall-CNN incorporates widely used kernels while ensuring transparency, making its decisions visually interpretable by physicians and clinicians. This balanced approach offers both the accuracy of deep learning models and the explainability needed for clinical applications. Full article

The weighted kappa coefficient is commonly used for assessing agreement between two raters on an ordinal scale. This study is the first to assess the impact of missing data on the value of weighted kappa. We compared three methods for handling missing data in a simulation study: predictive mean matching, listwise deletion and a weighted version of Gwet’s kappa. We compared their performances under three missing data mechanisms, using agreement tables with various numbers of categories and different values of weighted kappa. Predictive mean matching outperformed the other two methods in most simulated cases in terms of root mean squared error and in all cases in terms of bias. Full article

Artificial neural networks (ANNs) have been used for classification tasks involving functional magnetic resonance imaging (fMRI), though typically focusing only on fractions of the brain in the analysis. Recent work combined shallow neural networks (SNNs) with explainable artificial intelligence (xAI) techniques to extract insights into brain processes. While earlier studies validated this approach using motor task fMRI data, the present study applies it to Theory of Mind (ToM) cognitive tasks, using data from the Human Connectome Project’s (HCP) Young Adult database. Cognitive tasks are more challenging due to the brain’s non-linear functions. The HCP multimodal parcellation brain atlas segments the brain, guiding the training, pruning, and retraining of an SNN. Shapley values then explain the retrained network, with results compared to General Linear Model (GLM) analysis for validation. The initial network achieved 88.2% accuracy, dropped to 80.0% after pruning, and recovered to 84.7% post-retraining. SHAP explanations aligned with GLM findings and known ToM-related brain regions. This fMRI analysis successfully addressed a cognitively complex paradigm, demonstrating the potential of explainability techniques for understanding non-linear brain processes. The findings suggest that xAI, and knowledge extraction in particular, is valuable for advancing mental health research and brain state decoding. Full article

This work deals with the investigation and optimization of the MINDWALC node classification algorithm with a focus on its ability to learn human-interpretable decision trees from knowledge graph databases. For this, we introduce methods to optimize MINDWALC for a specific use case, in which the processed knowledge graph is strictly divided into its inner background knowledge (knowledge about a given domain) and instance knowledge (knowledge about given instances). We present the following improvement approaches, whereby the basic idea of MINDWALC—namely, to use discriminative walks through the knowledge graph as features—remains untouched. First, we apply relation-tail merging to give MINDWALC the ability to take relation-modified nodes into account. Second, we introduce walks with flexible walking depths, which can be used together with MINDWALC’s original walking strategy and can help to detect more similarities between node instances. In some cases, especially with hierarchical, incomplete tree-like structured graphs, our presented flexible walk can improve the classification performance of MINDWALC significantly. However, on mixed knowledge graph structures, the results are mixed. In summary, we were able to show that our proposed methods significantly optimize MINDWALC on tree-like structured graphs, and that MINDWALC is able to utilize background knowledge to replace missing instance knowledge in a human-comprehensible way. Our test results on our medical toy datasets indicate that our MINDWALC optimizations have the potential to enhance decision-making in medical diagnostics, particularly in domains requiring interpretable AI solutions. Full article

Generative large language models (LLMs) have revolutionized the development of knowledge-based systems, enabling new possibilities in applications like ChatGPT, Bing, and Gemini. Two key strategies for domain adaptation in these systems are Domain-Specific Fine-Tuning (DFT) and Retrieval-Augmented Generation (RAG). In this study, we evaluate the performance of RAG and DFT on several LLM architectures, including GPT-J-6B, OPT-6.7B, LLaMA, and LLaMA-2. We use the ROUGE, BLEU, and METEOR scores to evaluate the performance of the models. We also measure the performance of the models with our own designed cosine similarity-based Coverage Score (CS). Our results, based on experiments across multiple datasets, show that RAG-based systems consistently outperform those fine-tuned with DFT. Specifically, RAG models outperform DFT by an average of 17% in ROUGE, 13% in BLEU, and 36% in CS. At the same time, DFT achieves only a modest advantage in METEOR, suggesting slightly better creative capabilities. We also highlight the challenges of integrating RAG with DFT, as such integration can lead to performance degradation. Furthermore, we propose a simplified RAG-based architecture that maximizes efficiency and reduces hallucination, underscoring the advantages of RAG in building reliable, domain-adapted knowledge systems. Full article

Adversarial training, a widely used technique for fortifying the robustness of machine learning models, has seen its effectiveness further bolstered by modifying loss functions or incorporating additional terms into the training objective. While these adaptations are validated through empirical studies, they lack a solid theoretical basis to explain the models’ secure and robust behavior. In this paper, we investigate the integration of adversarial triplets within the adversarial training framework, a method previously shown to enhance robustness. However, the reasons behind this increased robustness are poorly understood, and the impact of different adversarial triplet configurations remains unclear. To address this gap, we utilize the robust and non-robust features framework to analyze how various adversarial triplet compositions influence robustness, providing deeper insights into the robustness guarantees of this approach. Specifically, we introduce a novel framework that explains how different compositions of adversarial triplets lead to distinct training dynamics, thereby affecting the model’s adversarial robustness. We validate our theoretical findings through empirical analysis, demonstrating that our framework accurately characterizes the effects of adversarial triplets on the training process. Our results offer a comprehensive explanation of how adversarial triplets influence the security and robustness of models, providing a theoretical foundation for methods that employ adversarial triplets to improve robustness. This research not only enhances our theoretical understanding but also has practical implications for developing more robust machine learning models. Full article

The integration of machine learning (ML) with big data has revolutionized industries by enabling the extraction of valuable insights from vast and complex datasets. This convergence has fueled advancements in various fields, leading to the development of sophisticated models capable of addressing complicated problems. However, the application of ML in big data environments presents significant challenges, including issues related to scalability, data quality, model interpretability, privacy, and the handling of diverse and high-velocity data. This survey provides a comprehensive overview of the current state of ML applications in big data, systematically identifying the key challenges and recent advancements in the field. By critically analyzing existing methodologies, this paper highlights the gaps in current research and proposes future directions for the development of scalable, interpretable, and privacy-preserving ML techniques. Additionally, this survey addresses the ethical and societal implications of ML in big data, emphasizing the need for responsible and equitable approaches to harnessing these technologies. The insights presented in this paper aim to guide future research and contribute to the ongoing discourse on the responsible integration of ML and big data. Full article

This study introduces the Pixel-Level Interpretability (PLI) model, a novel framework designed to address critical limitations in medical imaging diagnostics by enhancing model transparency and diagnostic accuracy. The primary objective is to evaluate PLI’s performance against Gradient-Weighted Class Activation Mapping (Grad-CAM) and achieve fine-grained interpretability and improved localization precision. The methodology leverages the VGG19 convolutional neural network architecture and utilizes three publicly available COVID-19 chest radiograph datasets, consisting of over 1000 labeled images, which were preprocessed through resizing, normalization, and augmentation to ensure robustness and generalizability. The experiments focused on key performance metrics, including interpretability, structural similarity (SSIM), diagnostic precision, mean squared error (MSE), and computational efficiency. The results demonstrate that PLI significantly outperforms Grad-CAM in all measured dimensions. PLI produced detailed pixel-level heatmaps with higher SSIM scores, reduced MSE, and faster inference times, showcasing its ability to provide granular insights into localized diagnostic features while maintaining computational efficiency. In contrast, Grad-CAM’s explanations often lack the granularity required for clinical reliability. By integrating fuzzy logic to enhance visual and numerical explanations, PLI can deliver interpretable outputs that align with clinical expectations, enabling practitioners to make informed decisions with higher confidence. This work establishes PLI as a robust tool for bridging gaps in AI model transparency and clinical usability. By addressing the challenges of interpretability and accuracy simultaneously, PLI contributes to advancing the integration of AI in healthcare and sets a foundation for broader applications in other high-stake domains. Full article

Basketball players are traditionally classified into five positions. This study examines the correlation between player performance, game statistics, and designated positions. It also explores how statistical contributions have evolved over time. Machine learning classifiers were used to identify key metrics that distinguish player positions and determine the most effective classification algorithms. Our findings confirm a correlation between game statistics and positions, reinforcing the relevance of traditional roles. However, results also show that modern players contribute across multiple positions, reflecting a shift toward versatility. Despite this flexibility, players maintain distinct roles and responsibilities. This classification approach highlights key performance metrics and lays the groundwork for future clustering and mapping analysis, offering deeper insights into player roles and team dynamics in contemporary basketball. Full article

Transformer models produce advanced text representations that have been used to break through the hard challenge of natural language understanding. Using the Transformer’s attention mechanism, which acts as a language learning memory, trained on tens of billions of words, a word sense disambiguation (WSD) algorithm can now construct a more faithful vectorial representation of the context of a word to be disambiguated. Working with a set of 34 lemmas of nouns, verbs, adjectives and adverbs selected from the National Reference Corpus of Romanian (CoRoLa), we show that using BERT’s attention heads at all hidden layers, we can devise contextual vectors of the target lemma that produce better clusters of lemma’s senses than the ones obtained with standard BERT embeddings. If we automatically translate the Romanian example sentences of the target lemma into English, we show that we can reliably infer the number of senses with which the target lemma appears in the CoRoLa. We also describe an unsupervised WSD algorithm that, using a Romanian BERT model and a few example sentences of the target lemma’s senses, can label the Romanian induced sense clusters with the appropriate sense labels, with an average accuracy of 64%. Full article

This study investigates the relationship between consumer personality traits, specifically openness, and responses to product designs. Consumers are categorized based on their levels of openness, and their affective responses to nine vase designs, varying in curvature and line quantity, are evaluated. The study then introduces the inverse clustering approach, which prioritizes maximizing predictive model accuracy over within-cluster similarity. This method iteratively refines cluster assignments to optimize prediction performance, minimizing errors in forecasting consumer design preferences. The results demonstrate that the inverse clustering approach yields more effective clusters than personality-based clustering. Moreover, while there is some overlap between personality-based and accuracy-based clustering, the inverse clustering method captures additional individual characteristics, extending beyond personality traits and improving the understanding of consumer product design response. The practical implications of this study are significant for product designers, as it enables the development of more personalized designs and optimization of product features to enhance specific consumer perceptions, such as robustness or esthetic appeal. Full article

Representing and interpreting human opinions within an unstructured framework is inherently complex. Hesitant fuzzy linguistic term sets offer a comprehensive context that facilitates a nuanced understanding of diverse perspectives. This study introduces a methodology that integrates sentiment analysis with hesitant fuzzy linguistic term sets to effectively aggregate and compare news from diverse sources. By employing linguistic scales, our approach enhances the interpretation of various perceptions and attitudes, facilitating comprehensive knowledge extraction and representation. The main objective of this research is to conduct a comparative analysis of news coverage across European countries in relation to the Israel–Gaza war. This analysis aims to capture the multifaceted sensitivities surrounding the ongoing situation, highlighting how different nations perceive the conflict. Full article

Interdisciplinary research (IDR) is essential for addressing complex global challenges that surpass the capabilities of any single discipline. However, measuring interdisciplinarity remains challenging due to conceptual ambiguities and inconsistent methodologies. To overcome these challenges, we propose a deep learning approach that quantifies interdisciplinarity in scientific articles through semantic analysis of titles and abstracts. Utilizing the Semantic Scholar Open Research Corpus (S2ORC), we leveraged metadata field tags to categorize papers as either interdisciplinary or monodisciplinary, establishing the foundation for supervised learning in our model. Specifically, we preprocessed the textual data and employed a Text Convolutional Neural Network (Text CNN) architecture to identify semantic patterns indicative of interdisciplinarity. Our model achieved an F1 score of 0.82, surpassing baseline machine learning models. By directly analyzing semantic content and incorporating metadata for training, our method addresses the limitations of previous approaches that rely solely on bibliometric features such as citations and co-authorship. Furthermore, our large-scale analysis of 136 million abstracts revealed that approximately 25% of the literature within the specified disciplines is interdisciplinary. Additionally, we outline how our quantification method can be integrated into a TRIZ-based (Theory of Inventive Problem Solving) methodological framework for cross-disciplinary innovation, providing a foundation for systematic knowledge transfer and inventive problem solving across domains. Overall, this approach not only offers a scalable measurement of interdisciplinarity but also contributes to a framework for facilitating innovation through structured cross-domain knowledge integration. Full article

Artificial intelligence (AI) is increasingly applied in a wide range of healthcare and Intensive Care Unit (ICU) areas to serve—among others—as a tool for disease detection and prediction, as well as for healthcare resources’ management. Since sepsis is a high mortality and rapidly developing organ dysfunction disease afflicting millions in ICUs and costing huge amounts to treat, the area can benefit from the use of AI tools for early and informed diagnosis and antibiotic administration. Additionally, resource allocation plays a crucial role when patient flow is increased, and resources are limited. At the same time, sensitive data use raises the need for ethical guidelines and reflective datasets. Additionally, explainable AI is applied to handle AI opaqueness. This study aims to present existing clinical approaches for infection assessment in terms of scoring systems and diagnostic biomarkers, along with their limitations, and an extensive overview of AI applications in healthcare and ICUs in terms of (a) sepsis detection/prediction and sepsis mortality prediction, (b) length of ICU/hospital stay prediction, and (c) ICU admission/hospitalization prediction after Emergency Department admission, each constituting an important factor towards either prompt interventions and improved patient wellbeing or efficient resource management. Challenges of AI applications in ICU are addressed, along with useful recommendations to mitigate them. Explainable AI applications in ICU are described, and their value in validating, and translating predictions in the clinical setting is highlighted. The most important findings and future directions including multimodal data use and Transformer-based models are discussed. The goal is to make research in AI advances in ICU and particularly sepsis prediction more accessible and provide useful directions on future work. Full article

Balancing the accuracy and the complexity of models is a well established and ongoing challenge. Models can be misleading if they are not accurate, but models may be incomprehensible if their accuracy depends upon their being complex. In this paper, semilattices are examined as an option for balancing the accuracy and the complexity of machine learning models. This is done with a type of machine learning that is based on semilattices: algebraic machine learning. Unlike trees, semilattices can include connections between elements that are in different hierarchies. Trees are a subclass of semilattices. Hence, semilattices have higher expressive potential than trees. The explanation provided here encompasses diagrammatic semilattices, algebraic semilattices, and interrelationships between them. Machine learning based on semilattices is explained with the practical example of urban food access landscapes, comprising food deserts, food oases, and food swamps. This explanation describes how to formulate an algebraic machine learning model. Overall, it is argued that semilattices are better for balancing the accuracy and complexity of models than trees, and it is explained how algebraic semilattices can be the basis for machine learning models. Full article

Urban happiness prediction presents a complex challenge, due to the nonlinear and multifaceted relationships among socio-economic, environmental, and infrastructural factors. This study introduces an advanced hybrid model combining a gradient boosting machine (GBM) and neural network (NN) to address these complexities. Unlike traditional approaches, this hybrid leverages a GBM to handle structured data features and an NN to extract deeper nonlinear relationships. The model was evaluated against various baseline machine learning and deep learning models, including a random forest, CNN, LSTM, CatBoost, and TabNet, using metrics such as RMSE, MAE, R², and MAPE. The GBM + NN hybrid achieved superior performance, with the lowest RMSE of 0.3332, an R² of 0.9673, and an MAPE of 7.0082%. The model also revealed significant insights into urban indicators, such as a 10% improvement in air quality correlating to a 5% increase in happiness. These findings underscore the potential of hybrid models in urban analytics, offering both predictive accuracy and actionable insights for urban planners. Full article

The now-globally recognized concerns of AI’s environmental implications resulted in a growing awareness of the need to reduce AI carbon footprints, as well as to carry out AI processes responsibly and in an environmentally friendly manner. Benchmarking, a critical step when evaluating AI solutions with machine learning models, particularly with language models, has recently become a focal point of research aimed at reducing AI carbon emissions. Contemporary approaches to AI model benchmarking, however, do not enforce (nor do they assume) a model initial selection process. Consequently, modern model benchmarking is no different from a “brute force” testing of all candidate models before the best-performing one could be deployed. Obviously, the latter approach is inefficient and environmentally harmful. To address the carbon footprint challenges associated with language model selection, this study presents an original benchmarking approach with a model initial selection on a proxy evaluative task. The proposed approach, referred to as Language Model-Dataset Fit (LMDFit) benchmarking, is devised to complement the standard model benchmarking process with a procedure that would eliminate underperforming models from computationally extensive and, therefore, environmentally unfriendly tests. The LMDFit approach draws parallels from the organizational personnel selection process, where job candidates are first evaluated by conducting a number of basic skill assessments before they would be hired, thus mitigating the consequences of hiring unfit candidates for the organization. LMDFit benchmarking compares candidate model performances on a target-task small dataset to disqualify less-relevant models from further testing. A semantic similarity assessment of random texts is used as the proxy task for the initial selection, and the approach is explicated in the context of various text classification assignments. Extensive experiments across eight text classification tasks (both single- and multi-class) from diverse domains are conducted with seven popular pre-trained language models (both general-purpose and domain-specific). The results obtained demonstrate the efficiency of the proposed LMDFit approach in terms of the overall benchmarking time as well as estimated emissions (a 37% reduction, on average) in comparison to the conventional benchmarking process. Full article

Data drift can significantly impact the outcome of a model. Early detection of data drift is crucial for ensuring user confidence in predictions. It allows the user to check if a particular model needs retraining using updated data to adapt to the evolving process dynamics. This study compares five different statistical tests, namely four unidimensional and a new multidimensional test (MSPC), to identify data drift in both mean and deviation. While some are designed to detect drift in mean only, like our multidimensional proposal, others respond to changes in both mean and deviation. However, our Hotelling multidimensional method can be trained once and then applied in a single stage to any data stream with several attributes, and it can identify the most relevant variables causing a data drift with one execution, thus avoiding the need for a single univariate test for each attribute. Moreover, our method yields the relative importance of each attribute for drift and allows users to increase or decrease the relative weight of each variable regarding drift detection. It also may be capable of detecting drift due to changes in multivariate interactions. This behavior is especially suitable for real-world scenarios, such as industry, finance, or healthcare environments. Full article

Deep learning models are widely used for medical image analysis and require large datasets, while sufficient high-quality medical data for training are scarce. Data augmentation has been used to improve the performance of these models. The lack of transparency of complex deep-learning models raises ethical and judicial concerns inducing a lack of trust by both medical experts and patients. In this paper, we focus on evaluating the impact of different data augmentation methods on the explainability of deep learning models used for medical image classification. We investigated the performance of different traditional, mixing-based, and search-based data augmentation techniques with DenseNet121 trained on chest X-ray datasets. We evaluated how the explainability of the model through correctness and coherence can be impacted by these data augmentation techniques. Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) methods were used. Sanity checks and overlapping scores were applied to confirm the correctness and coherence of explainability. The results indicate that both LIME and SHAP passed the sanity check regardless of the type of data augmentation method used. Overall, TrivialAugment performs the best on completeness and coherence. Flipping + cropping performs better on coherence using LIME. Generally, the overlapping scores for SHAP were lower than those for LIME, indicating that LIME has a better performance in terms of coherence. Full article