Educational Data Mining (EDM) increasingly depends on large, high-quality datasets to drive predictive and adaptive learning systems. However, data scarcity, privacy restrictions, and limited accessibility severely hinder research reproducibility and cross-institutional collaboration. Synthetic data generation provides an emerging solution, enabling the creation of artificial yet statistically realistic datasets that preserve analytical utility while preserving student privacy. This study benchmarks four generative approaches, namely Gaussian Copula, CopulaGAN, Conditional Tabular Generative Adversarial Networks (CTGAN), and Tabular Variational Auto Encoders (TVAE), on student data from six undergraduate courses at a European university. Using the open-source Synthetic Data Vault (SDV) framework, we evaluate the fidelity and Machine Learning utility of synthetic student records through Random Forest classifiers across five metrics, namely accuracy, F1-score, precision, recall, and Area Under Curve (AUC). The results show that synthetic data can achieve 96-98% of the predictive performance obtained when training on real data, with TVAE consistently demonstrating the highest multivariate fidelity. Our contributions are threefold: (i) we introduce a reproducible benchmarking pipeline for synthetic data evaluation in educational settings; (ii) we empirically compare statistical and deep generative synthesizers on real-world tabular student data; and (iii) we identify critical research directions related to privacy and reproducibility. The findings position synthetic data generation as a foundational technology for ethical and privacy-preserving EDM.

Topological data analysis (TDA) has evolved into a flexible and robust paradigm for obtaining qualitative, geometry-inspired insights from high-dimensional, noisy, and complex data. Grounded in algebraic topology, geometry, statistics, and machine learning (ML), TDA provides multiscale descriptions through persistent homology, Mapper (a graph-based method that summarizes the shape of high-dimensional data), and related topological signatures that are often inaccessible to standard linear and metric methods. In recent years, and especially during 2024–2025, TDA has expanded rapidly across science, engineering, biomedical research, and socio-economic studies, while also being integrated with modern learning paradigms such as deep learning (DL) and graph learning. This survey summarizes recent developments in TDA using a carefully selected set of articles, with emphasis on 2024–2025. We first present the mathematical and computational foundations of TDA, covering simplicial complexes, filtrations, persistent homology, the Mapper algorithm, and computational advances such as data simplification, stability, and efficiency. We then review applications in time series and dynamical systems, biomedical imaging and precision medicine, engineering and physical sciences, finance and risk analysis, DL and interpretability, and security and critical infrastructure systems. Throughout, we highlight how TDA can extract informative features, function as a model component, and provide a conceptual lens for studying complex systems. However, the survey also emphasizes recurrent failure patterns: TDA performance is highly sensitive to filtration, embedding, and vectorization choices; aggressive simplification can dilute or remove informative topological signals; and integration into standard ML workflows still lacks uniform validation and reporting protocols. We conclude by outlining key challenges—including scalability, statistical foundations, interpretability, and compatibility with rapidly evolving artificial intelligence (AI) paradigms—and by identifying directions for future research. The survey also provides a unifying design perspective for TDA systems, highlighting methodological trade-offs and emerging research directions for integrating topology with modern ML.

The rapid expansion of global e-commerce platforms has led to unprecedented volumes of heterogeneous, multimodal, and continuously evolving data, creating significant challenges for prediction, personalization, trust, and operational decision-making. Deep Learning has emerged as a core enabling technology for addressing these challenges, offering powerful representation learning, sequential reasoning, graph-based inference, and decision-centric optimization capabilities. This survey provides a comprehensive and decision-oriented review of recent advances in Deep Learning for e-commerce, covering consumer behavior prediction, demand forecasting, recommendation systems, sentiment and review intelligence, catalogue understanding, fraud detection, cybersecurity, and large-scale operational optimization. Beyond predictive and personalization tasks, the survey emphasizes decision intelligence, highlighting the growing role of Reinforcement Learning and integrated Artificial Intelligence systems in pricing, logistics, warehouse automation, and platform reliability. We organize the literature according to key e-commerce objectives and operational contexts, analyze methodological trends and deployment challenges, and discuss limitations related to scalability, robustness, interpretability, and cross-border adaptability. Finally, we identify open research directions toward unified multimodal foundation models, culturally adaptive intelligence, and trustworthy, sustainable Artificial Intelligence systems for next-generation e-commerce platforms.

Multimodal Learning Analytics (MMLA) is an extension of Learning Analytics that combines multiple data streams such as audio, video, physiological signals, logs, and spatial trails to analyze learning processes that cannot be easily captured through any single modality. This review synthesizes research on sensing and instrumentation, feature extraction, multimodal fusion, modeling approaches, and end-to-end systems that provide feedback and support reflection. We also discuss how generative AI and Large Language Models (LLMs) increasingly improve MMLA pipelines by enabling scalable semantic and pragmatic analysis of learner discourse and interaction. In addition, we review robustness issues that arise when working with real-world data (e.g., noise, missing data, and scalability) and responsible deployment issues such as privacy and student-focused views of fairness, accountability, transparency, and ethics (FATE).

Tabular classification is a central supervised learning problem in high-stakes domains. While tree ensembles are strong baselines and deep tabular models sometimes help, standard ensembles typically use global fusion (e.g., uniform averaging), implicitly assuming constant expert reliability across the input space. We propose a lightweight, architecture agnostic dynamic gating mechanism that performs instance-wise mixture weighting using raw covariates, expert probability vectors, and two confidence features (entropy and margin). Across 14 benchmark datasets, dynamic gating improves mean-ensemble accuracy from 0.9047 to 0.9149 in a heterogeneous pool ( +0.0102 ) and from 0.9082 to 0.9126 in a tree-only pool ( +0.0044 ), with 11/14 dataset wins in both settings. Aggregate paired Wilcoxon tests confirm significance ( p=2.44× 10^-4 for heterogeneous, p=1.07× 10^-2 for tree-only), showing that compact instance-wise weighting improves robustness without architecture-specific fusion.

Supervised fault diagnosis of inverter-driven permanent magnet synchronous motors (PMSMs) typically requires large volumes of expert-annotated data, which are expensive to obtain in industrial settings. This paper presents a systematic benchmark of six pool-based active learning (AL) strategies, random sampling, uncertainty (entropy), Bayesian Active Learning by Disagreement (BALD), Query-by-Committee (vote entropy), CoreSet (k-center), and a hybrid uncertainty–diversity method, combined with four tree-based classifiers on a nine-class PMSM fault dataset comprising 10 857 samples and 25 engineered features. Experiments are conducted under stratified five-fold cross-validation across ten labelling budgets (1–100 0.9998 ± 0.0004 using only 10

Visual Language Models (VLMs) have swiftly accelerated the blending of the visual modality with textual information, enabling more natural and contextually aware human–AI interaction. This survey offers an in-depth overview of foundations, capabilities and constraints of existing VLMs.We first introduce a taxonomy of architecture, objectives of training and benchmark data, commenting on how design decisions impact performance. Then, we discuss evaluation processes, balancing capability measures with reliability measures like robustness, fairness and efficiency. Applications across fields like healthcare, education, scientific discovery and creative arts are discussed to illuminate both potential and practical challenges. Finally, we summarize open issues in scalability, explainability and trustworthy operation. Aiming to marry technical expertise with applied direction, the present survey hopes to illuminate researchers, practitioners and educators with an appreciation of the present state of VLMs and their future responsible use.

Tabular classification remains a core workload in applied machine learning, yet ensemble integration in practice is often reduced to uniform probability averaging. This paper provides a focused comparison of three classical integration strategies over tabular expert pools, Mean Ensemble, Static Weighted Ensemble, and Stacking. We evaluate 14 datasets under two expert regimes, tree-only and hybrid. Across datasets, aggregated accuracy improves from 0.9064 ± 0.0791 (Mean Ensemble) to 0.9133 ± 0.0754 (Static Weighted Ensemble) and 0.9158 ± 0.0761 (Stacking). The average gains versus mean are +0.0069 ± 0.0080 for static weighting and +0.0094 ± 0.0113 for stacking. Wilcoxon signed-rank tests show significant improvements over mean in the aggregated view (static, p=0.008775 ; stacking, p=0.001871 ), with strong evidence in the hybrid pool. Stacking is top-ranked most often in 9/14 datasets, compared with static weighting in 3/14 datasets and mean averaging being top-ranked or tied in 2/14. Overall, classical stacking is a robust, low-complexity upgrade over uniform averaging, especially when expert diversity is higher.

This study investigates how teacher-education undergraduates in Greece use, evaluate, and trust Artificial Intelligence (AI) in higher education, with particular attention to the gap between widespread adoption and limited epistemic trust. The topic is important because generative AI is rapidly entering universities, reshaping learning practices, academic integrity, and the legitimacy of knowledge, while learners often rely on systems whose outputs are not easily verifiable. The study focuses on future teachers because they are both current users of AI in higher education and likely future mediators of its use in school settings. Addressing this problem, the study contributes empirical evidence on how AI adoption relates to epistemic authority and institutional legitimacy within teacher education rather than across university students in general. A mixed-methods design was employed using a structured questionnaire completed by 363 teacher-education undergraduates from the University of Patras and the University of Ioannina in Greece; the sample was predominantly women (86.0%) and first-year students (92.6%). Quantitative responses were analyzed statistically, open-ended answers were examined thematically, and factor analysis was used to identify latent attitudinal dimensions. The findings indicate very high AI use in everyday life (92.6%) and study practices (81.3%), but only moderate trust: 1.4% reported complete trust and 12.1% generally trusted AI-generated answers. Six dimensions explained 61.73% of total variance, pointing to a layered attitudinal structure within this teacher-education population, consistent with an adoption-trust paradox and with the need for transparent, verifiable, human-supervised educational AI. The observed verification-based trust calibration may partly reflect an emerging pedagogical orientation toward source checking and responsibility for knowledge mediation, but given the strong concentration of first-year students, this should be interpreted as characteristic of early-stage teacher education rather than of university students more broadly.

We examine whether pool-based active learning can improve predictive performance without worsening group fairness under scarce labels. We compare a fairness-only baseline with three lightweight query strategies on two gender-sensitive learning-analytics tasks: MBA admissions and OULAD student success prediction. The strategies are random sampling, uncertainty entropy, and uncertainty margin. The pipeline evaluates decision trees and random forests under label budgets from 20

Class imbalance degrades rare-class ranking in tabular classification, while conventional oversampling trains the whole classifier under a synthetic balanced view of the data. This paper studies Prior–Shadow Tree Ensembles (PSTE), an alternative for imbalanced tabular tree ensembles which retains an original-prior branch, adds an oversampled shadow branch, and blends the two under calibration and precision constraints rather than relying on a single oversampled model. The resulting adaptive PSTE framework selects the shadow contribution and probability-correction by nested validation while keeping the tree budget fixed. The central hypothesis is that rare class ranking can be improved without fully replacing the empirical training prior. We also examine Fast Outward Ladder, a geometry guided extension of SMOTE-style interpolation, to assess whether outward minority expansion provides more useful support for tree models than vanilla SMOTE. We evaluate the approach on 25 imbalanced tabular benchmarks and three tree backbones against classical oversampling, cleaning, undersampling, and imbalanced-ensemble baselines, using PR-AUC as the primary endpoint. Across the resulting 75 dataset–backbone model blocks, the Friedman test rejects equal PR-AUC ranks (𝜒f2 = 544.90, 𝑝 = 1.54 × 10−100), and every PSTE-wrapped sampler ranks above its sampler-only counterpart. PSTE–Fast Outward Ladder attains the highest overall mean PR-AUC, 0.6870, while its logit variant gives the best average rank. Against sampler-only SMOTE, PSTE–Fast Outward Ladder improves mean and median PR-AUC by 0.0204 and 0.0121, wins 68 of 75 matched blocks, and remains significant after Holm correction (𝑝 < 0.001). At the sampler-only level, Fast Outward Ladder also improves over SMOTE by +0.0090 mean PR-AUC and wins 56 of 75 blocks, indicating that the sampler itself contributes beyond the PSTE wrapper.

Imbalanced tabular classification remains a central challenge in clinical risk modeling, where minority-class misses can be costly and subgroup performance disparities are undesirable. This study evaluates augmentation for diabetes-focused classification with six classical oversamplers (SMOTE, ADASYN, BORDERLINE_SMOTE, SVM_SMOTE, KMEANS_SMOTE, SMOTE_BORDERLINE2) and four generative baselines (Copula, TVAE, CTGAN, GREAT). The four generative baselines were executed within the MAGA control workflow, whereas the classical oversamplers were retained as direct end-to-end baselines. We study MAGA (Minority class Augmentation with Generative Approach) as a control framework around minority-focused generative augmentation rather than as a new generator. MAGA links four operational controls: candidate generation, fidelity-based acceptance filtering, adaptive top-up toward target coverage, and fairness-aware model/threshold selection via Composite Validation Utility (CVU), denoted as JVAL in pipeline outputs. The benchmark covers two datasets (OpenML diabetes and an NHANES-derived diabetes subset), three target imbalance levels (60%, 80%, 100% minority-to-majority), two seeds, and six classifiers. Across 720 matched combinations, GREAT utilizing the MAGA framework attains the highest mean test PR-AUC (0.6374) and mean test recall (0.9063), while SMOTE leads on mean JVAL (0.6958). At the configuration level, the best observed PR-AUC is obtained by GREAT on OpenML diabetes at the 100% target ratio with HistGB (0.7780), exceeding the best SMOTE result on the same dataset and ratio (0.7592 with LogReg). Holm-corrected pairwise tests show that GREAT significantly outperforms CTGAN, ADASYN, COPULA, BORDERLINE_SMOTE, SVM_SMOTE, and TVAE on PR-AUC, but not the strongest classical baselines. A post-hoc fairness audit using both equal-opportunity disparity and demographic-parity disparity shows that no single method dominates all objectives. Overall, MAGA’s contribution is procedural, providing a reproducible way to connect candidate generation, plausibility control, optional top-up, and fairness-aware operating-point selection in one benchmarked workflow.

Financial fraud has become a critical challenge for modern financial systems due to the rapid growth of digital transactions, online banking services, and electronic payment platforms. Traditional rule-based fraud detection systems are increasingly inadequate in addressing the evolving and adaptive strategies employed by fraudsters. Consequently, Machine Learning (ML) and Deep Learning (DL) techniques have emerged as powerful tools for detecting fraudulent activities in large-scale financial datasets. This paper presents a comprehensive survey of ML/DL approaches for financial fraud detection. The survey systematically reviews existing research across multiple methodological paradigms, including classical supervised learning, anomaly detection, graph-based methods, deep neural networks, multimodal architectures, and cost-sensitive learning frameworks. Particular emphasis is placed on emerging techniques such as graph neural networks, transformer-based architectures, and federated learning approaches designed to address privacy and scalability challenges. In addition to reviewing model architectures, this work analyzes key challenges inherent to fraud detection systems, including extreme class imbalance, concept drift, adversarial behavior, data privacy constraints, and real-time deployment requirements. Furthermore, the survey examines evaluation methodologies, highlighting the limitations of commonly used metrics and discussing more realistic evaluation strategies that incorporate operational costs and risk management considerations. This paper also provides a structured taxonomy of fraud detection methods, comparative analyses of commonly used datasets, and a synthesis of current research trends. Finally, open challenges and promising research directions are identified, including adaptive learning systems, interpretable Artificial Intelligence models, graph-based behavioral modeling, and privacy-preserving collaborative fraud detection frameworks.

Privacy-preserving data mining (PPDM) seeks to extract useful patterns from shared data without revealing sensitive information. Within PPDM, knowledge hiding—encompassing both association rule hiding (ARH) and frequent itemset hiding (FIH)—forms a coherent family of techniques that sanitize transactional databases before release. This focused survey synthesizes the main algorithmic paradigms for knowledge hiding (1999–2026), covering heuristic sanitization, border-based and exact optimization via integer linear programming, constraint-based and graph-based formulations, emerging learning-guided support mechanisms, and extensions to utility mining and non-relational structures. We use a PRISMA-style search and selection protocol to make the evidence base transparent and to mitigate selection bias. We trace the evolution from early disclosure-limitation heuristics to graph-guided and knowledge-graph approaches, and we treat deep-learning, GNN, and federated graph-learning work as adjacent tools that may support candidate selection, representation learning, or distributed deployment rather than as replacements for classical hiding validation. We identify persistent challenges around scalability, infeasibility in LP formulations, and evaluation standardization, and outline directions for future research. Unlike broader PPDM overviews, this review centers exclusively on transactional knowledge hiding. Beyond cataloging algorithms, it compares method families through their intervention mechanisms, side-effect profiles, scalability assumptions, and benchmark regimes, and it distills reporting recommendations for more reproducible empirical evaluation.

This paper presents a unified benchmark of fairness-aware SSL for learning analytics on MBA and OULAD. We compare three base classifiers (Decision Tree, Random Forest, XGBoost), ten SSL wrappers, and a fairness-only baseline across four label rates. Fairness for gender is evaluated with demographic parity difference (DP), equalized odds difference (EO), and equal opportunity difference (EOp). Across 792 runs, the strongest utility-oriented selections are usually Self-Training, SETRED, and CoTrainingByCommittee, whereas the fairest outcomes are often obtained by DeTriTraining at a substantial accuracy cost. Overall, the results show a persistent but context-dependent utility–fairness trade-off across datasets and supervision regimes, implying that early-warning deployment requires explicit utility and disparity guardrails rather than a single globally best method.

The integration of human judgment into artificial intelligence (AI) systems has emerged as a key research direction, particularly for high-stakes applications where full automation remains insufficient. Human-in-the-Loop (HITL) AI represents a field that combines machine learning capabilities with human oversight, feedback, and decision-making at various stages of the AI pipeline. This survey provides a systematic review of HITL approaches, covering theoretical foundations, technical methods, ethical considerations, and domain-specific applications. We propose a unified taxonomy that categorizes HITL systems based on loop placement, interaction granularity, and temporal characteristics. This review synthesizes findings from healthcare, autonomous systems, cybersecurity, and other high-risk domains where human oversight is essential. We also examine the challenges of scalability, cognitive load, and trust calibration that affect the practical deployment of HITL systems. The final section outlines open research directions and introduces a framework for designing effective human–AI collaborative systems.

Artificial Intelligence (AI) has transitioned from a specialized research area to a ubiquitous socio-technical infrastructure influencing sectors from healthcare and law to manufacturing and defense. In tandem with its transformative promise, AI has created an exponentially expanding ethics literature questioning, fairness, transparency, accountability, and justice. This review synthesizes publications and key policy developments between 2019 and 2025, bringing sectoral discourses together with cross-cutting frameworks. Grounded in a systematic scoping review methodology, we frame the field along four meta-dimensions: trust and transparency, bias and fairness, governance & regulation, and justice, while we investigate their expression across diverse sectors. Special attention is dedicated to healthcare (patient trust and algorithmic bias), education (integrity and authorship), media (misinformation), law (accountability), and the industrial sector (data integrity, intellectual property protection, and environmental safety). We ground abstract principles in concrete case studies to illustrate real-world harms and mitigation strategies. Furthermore, we incorporate pluralistic ethics (e.g., Ubuntu, Islamic perspectives), environmental ethics, and emerging challenges posed by Generative AI and neuro-AI interfaces. To bridge theory and practice, we propose an operational governance framework for organizations. We contend that success involves transitioning from principles toward ethics-by-design, pluralistic governance, sustainability, and adaptive oversight. This review is intended for scholars, practitioners, and policymakers who need a comprehensive and actionable framework for navigating the complex landscape of AI ethics.

It is a fact that natural language processing (NLP) has become an integral part of daily life, with research outcomes being integrated into various everyday implementations. A significant portion of this success can reasonably be attributed to the architecture of transformers. In this context, text classification problems constitute a large part of ongoing research. Simultaneously, there is a growing demand for high-quality labeled textual data. The latter is becoming increasingly urgent with the rising complexity and size of models. Based on this, the present work investigates the integration of active learning strategies into text classification problems using transformer-based models from the BERT family. Through an extensive experimental framework involving 10 datasets and 7 different BERT-based classifiers, we demonstrate that the incorporation of active learning in the context of text classification can significantly reduce the need for labeled data during the fine-tuning procedures. Specifically, our experimental results illustrate that without sacrificing model effectiveness–as measured by various evaluation metrics–we can achieve at least a 50

Background: Multiple Sclerosis (MS) is a chronic autoimmune disease of the central nervous system with a propensity to inflict severe neurological disability. Accurate and early prediction of MS progression is extremely crucial for its management and treatment. Methods: In this paper, we compare a number of self-labeled semi-supervised learning methods used to predict MS from labeled and unlabeled medical data. Specifically, we compare the performance of Self-Training, SETRED, Co-Training, Co-Training by Committee, Democratic Co-Learning, RASCO, RelRASCO, CoForest, and TriTraining in different labeled ratios. The data contain clinical, imaging, and demographic features, allowing for a detailed comparison of each method's predictive ability. Results and Conclusions: The experimental results demonstrate that several self-labeling semi-supervised learning (SSL) algorithms perform competitively in the task of Multiple Sclerosis (MS) prediction, even when trained on as little as 30-40% of the labeled data. Notably, Co-Training by Committee, CoForest, and TriTraining consistently deliver high performance across all metrics (accuracy, F1-score, and MCC).

Fairness is a fundamental virtue in machine learning systems, alongside with four other critical virtues: Accountability, Transparency, Ethics, and Performance (FATE + Performance). Ensuring fairness has been a central research focus, leading to the development of various mitigation strategies in the literature. These approaches can generally be categorized into three main techniques: pre-processing (modifying data before training), in-processing (incorporating fairness constraints during training), and post-processing (adjusting outputs after model training). Beyond these, an increasingly explored avenue is the direct modification of existing algorithms, aiming to embed fairness constraints into their design while preserving or even enhancing predictive performance. This paper presents a comprehensive survey of classical machine learning models that have been modified or enhanced to improve fairness concerning sensitive attributes (e.g., gender, race). We analyze these adaptations in terms of their methodological adjustments, impact on algorithmic bias and ability to maintain predictive performance comparable to the original models.