Depression among university students is a critical global problem that affects academic performance, well-being, and dropout rates. Although machine learning models offer strong predictive power, their opacity limits their adoption in educational and clinical contexts. Most existing approaches emphasize either global or individual explanations, leaving a gap in multi-level interpretability that can inform institutional policy and personalized support. This paper introduces a novel multi-level explainable AI (XAI) framework that unifies SHAP for global importance ranking, H-LIME for hierarchical subgroup explanations, and counterfactual reasoning for actionable individual recourse. By establishing a generalizable theory of hierarchical interpretability that bridges global, group, and local levels of explanation, the framework advances both methodological rigor and practical relevance. Using a dataset of 27,901 students with demographic, academic, lifestyle and psychosocial factors, Random Forest achieved a strong predictive performance (accuracy = 0.839, ROC–AUC = 0.916). The results consistently identified suicidal ideation, academic pressure, and financial stress as dominant factors, while subgroup and individual analyses revealed contextual and personalized nuances that global explanations alone obscure. Individual-level counterfactuals further suggested feasible workload or lifestyle adjustments, offering practical recourse for at-risk students. Limitations include dependence on self-reported survey data and the fact that the sample is from a single country, which may limit the generalizability of the findings. This unified approach offers a reusable methodological framework for achieving multi-level interpretability in sensitive AI applications. It also provides evidence-based guidance for the early identification of at-risk individuals, customized interventions, and the development of equity-based institutional policies.

Recommending exercises in educational contexts requires balancing relevance and diversity to support effective learning progression. In such settings, content-based recommendation is particularly suitable, as it aligns with specific learning objectives and supports progression without relying on extensive user interaction history. This study introduces a content-based recommendation system (CBRS) designed to suggest programming exercises based on intrinsic characteristics of source code. The system evaluates syntactic, structural, statistical, complexity, and semantic attributes to provide a comprehensive assessment of exercise similarity. The approach addresses the challenge of identifying exercises that require similar knowledge or complexity levels to those previously completed by students. This capability is especially beneficial when students struggle with specific concepts, as it enables the recommendation of additional exercises targeting the same content to reinforce understanding and promote mastery. Each feature was first evaluated individually to determine its contribution to recommendation effectiveness. Selected combinations were then tested to examine how different attributes influence recommendation quality in terms of relevance and diversity. The full integration of all features resulted in a precision of 0.70, recall of 0.77, MRR of 0.64, coverage of 0.63, serendipity of 0.49, and novelty of 0.51. The proposed CBRS demonstrates strong alignment with learners’ exercise histories and the potential to generate recommendations that balance familiarity with exploration. Compared with collaborative filtering models on the same dataset, the CBRS achieved stronger performance in diversity-focused metrics, supporting more engaging and meaningful learning experiences.

Programming is an important skill for advancement in computer science and information and communication technology (ICT). Online judge (OJ) systems are increasingly popular platforms in programming education to enhance users programming skills through practice. However, in most OJ systems, the problems are usually listed in a simple format without clear identification by difficulty levels. This research aims to cluster problem difficulty effectively and analyze user performance. We extract features from the Aizu OJ (AOJ) submission log data and estimate user ability and problem difficulty scores using logistic regression (LR). We propose an approach that combines an attention mechanism (AM) with K-means applied to statistically selected features via pairwise correlation to categorize the problems. Compared to standard K-means and Fuzzy C-means (FCM) with their AM integration on full and reduced feature sets by principal component analysis (PCA), autoencoder (AE), and variance inflation factor (VIF), the proposed method demonstrates the best clustering performance. It achieves the highest silhouette score (SS) of 0.507, Calinski-Harabasz index (CHI) of 6581.088, and a comparatively good Davis Bouldin index (DBI) of 0.625.We found a strong correlation between user ability and problem-solving rates; beginners and intermediate users consistently struggle with solving more complex problems. To support this, web-based clustering and group-based recommendations have been developed that visualize problem difficulty clusters, enabling users to understand difficulty levels and select problems based on their ability level. Moreover, the proposed approach is promising for effective clustering of more complex educational datasets and improving learning outcomes in programming education.

This study introduces and compares three innovative approaches for recommending programming problems within an Online Judge system (OJ), tackling the challenge of deriving implicit ratings from user interactions without explicit user ratings. Conventional collaborative filtering (CF) methods often struggle with the sparse and implicit feedback typical in educational settings, limiting recommendation accuracy. A common method to handle implicit feedback is to use a binary system, which in the context of programming problems, simply records whether a user solved a problem (1) or did not (0). However, this method lacks scalability, fails to capture satisfaction levels, and overlooks problem difficulty and user skill. We analyzed submissions on the OJ to develop methods that generate implicit ratings reflecting user engagement and problem-solving performance. We proposed three distinct approaches, i.e., a Direct Weighted Sum Approach, a Formula-Based Approach and a Distance-Based Approach. Each approach determines user satisfaction, scaled from 1 to 5, and utilizes this in a Neural CF model to improve recommendation accuracy. Additionally, we included a baseline dataset that categorizes exercises in a binary format, marking them as either solved or not. The results demonstrate that all the approaches performed well in terms of accuracy. However, the proposed approaches significantly outperformed the baseline approach in recommendation quality by presenting relevant items to users earlier and offering a broader range of items. Another advantage of the proposed approaches is their consideration of problem difficulty and user skill, enabling a better understanding of how challenging each recommended item might be for each user.

The evaluation of robotic systems has traditionally centered on task performance metrics such as accuracy, efficiency, and execution time. However, as robotics continues to expand across diverse domains, a broader and more structured assessment framework is necessary. This survey examines metrics from 113 research papers published between 2020 and 2024 to identify trends in robotic evaluation. The findings reveal a wide variety of metrics applied based on different research objectives, making standardization difficult. Despite this diversity, these metrics can generally be classified into task performance and non-task performance categories. Many studies prioritize task performance, often placing less emphasis on software contributions, which are frequently underrepresented in evaluations despite their integral role in control and decision-making. Additionally, robotic adaptability in harsh environments is often inconsistently assessed, limiting comparability across different studies. To address these challenges, this paper advocates for clearer frameworks that systematically integrate both hardware and software contributions. Furthermore, a standardized approach to evaluating robotic adaptability across various environmental conditions is recommended. Establishing structured methodologies will enhance comparability and lead to more precise assessments of robotic performance.