explainable-ai
Filtering by topic explainable-ai(3)Clear all filters
- PaperComputers & Education14 Jul 2026
LLM-derived metrics in second language writing assessment: an explainable AI approach
Jingying Hu, Yan Cong
This study derived surprisal, perplexity, and embedding-based similarity from large language models to quantify linguistic predictability and semantic coherence in second language writing. These metrics, computed on Chinese learner essays, decreased with proficiency for Traditional Chinese-focused models, while embedding similarity increased. Combining LLM-derived metrics with classical linguistic features improved proficiency classification, supporting transparent and scalable AI assessment for underrepresented languages.
Original abstract
Large language models (LLMs) are increasingly used in educational technology for automated writing assessment, yet most applications rely on prompt-based scoring and feedback generation, which often lack transparency, reproducibility, and interpretability. This study investigates whether model-internal LLM representations can provide interpretable and reproducible metrics for second language (L2) writing assessment. We derived surprisal and perplexity from next-token prediction to quantify linguistic predictability and embedding-based similarity to measure semantic coherence across sentences. These metrics were computed at the token, sentence, and discourse levels using three pretrained Chinese language models and evaluated on 1,196 essays written by Chinese L2 learners across four proficiency levels. Their relationships with 11 established linguistic measures of fluency, lexical sophistication, phraseological complexity, and syntactic complexity were also examined. Results showed that surprisal and perplexity generally decreased with proficiency for the two Traditional Chinese-focused models, indicating greater linguistic predictability in more proficient writing, whereas the multilingual model showed weaker sensitivity. Embedding-based similarity increased with proficiency, reflecting stronger semantic coherence. Combining LLM-derived metrics with classical linguistic features improved proficiency classification and prediction beyond either feature set alone. Correlation and qualitative analyses further demonstrated that the proposed metrics capture complementary aspects of writing while revealing conditions under which their interpretations become less reliable. These findings demonstrate the value of interpretable, model-derived metrics for transparent, reproducible, and scalable AI-supported L2 writing assessment, particularly for underrepresented learner populations and lower-resource languages.
- PaperComputers and Education: Artificial Intelligence11 Jul 2026
Fair and explainable educational recommendations with a hybrid Graph-GRU framework
Edmund Evangelista, Syed M. Salman Bukhari
The study introduces a Hybrid Heterogeneous Knowledge Graph-Gated Recurrent Unit framework that combines graph embeddings with sequential modeling to provide fair, robust, diverse, and explainable educational recommendations. It integrates multi-objective training, reranking, and model-centric explainability, achieving strong predictive performance on a Moodle LMS dataset while improving intra-list diversity and moderate catalogue coverage. However, popularity bias remains evident, highlighting the need for continued attention to fairness.
Original abstract
Artificial Intelligence (AI) recommender systems are increasingly used in education to personalize learning and help students navigate large collections of digital learning resources. However, many existing approaches emphasize predictive accuracy over fairness, robustness, diversity, and transparency. This creates an important educational challenge. The students with limited participation histories may receive less reliable support, while highly popular resources may dominate recommendation lists and limit access to other useful learning materials. To address this challenge, this study aims to develop and evaluate a responsible educational recommender framework that supports personalized learning resource navigation while making recommendation behaviour more fair, stable, diverse, and explainable. This study introduces the Hybrid Heterogeneous Knowledge Graph-Gated Recurrent Unit (Hybrid HKG-GRU) framework, which combines heterogeneous graph embeddings with sequential modelling to capture both the relational structure of course materials and the temporal dynamics of learner interactions. The framework integrates three contributions: (i) multi-objective training with Group Distributionally Robust Optimization (GroupDRO), (ii) Maximum Marginal Relevance (MMR) reranking to reshape exposure patterns, and (iii) built-in, model-centric explainability through path-based and counterfactual analyses. The empirical evaluation shows strong predictive performance with HR@10 = 0.68 and MRR = 0.41 on Moodle LMS logs dataset that comprises of 152 students, 59 resources, and approximately 150k interactions. The model also achieves high intra-list diversity and moderate catalogue coverage, while showing moderate counterfactual stability for many learners (median CR@10 = 1.0), although catalogue-level popularity bias remains evident. The framework provides model-centric interpretability and verification intended to support more transparent educational recommendation. The study positions the framework as a technically auditable approach for improving how learning resources are recommended, inspected, and monitored in educational settings. By integrating fairness, robustness, and explainability as coequal design objectives, the Hybrid HKG-GRU provides a methodological foundation for more responsible and accountable recommender systems in education and related high-stakes contexts.
- PaperComputers and Education: Artificial Intelligence19 Jun 2026
Federated and explainable learning analytics for privacy-preserving academic risk modeling across heterogeneous educational institutions
William Villegas-Ch, Alexandra Maldonado Navarro, Jaime Govea, Joselin García-Ortiz et al.
This study proposes a federated, explainable learning analytics framework for privacy-preserving academic risk modeling across heterogeneous institutions. The framework integrates temporal behavioral and socio-academic features in a multitask learning scheme. Findings show that federated models maintain strong discriminative performance despite heterogeneity, but calibration metrics are more sensitive to distributional shifts.
Original abstract
The increasing digitization of higher education has enabled the development of learning analytics models to identify students at risk of academic failure or dropout; however, most existing approaches rely on centralized training and assume homogeneous data distributions, limiting their applicability across institutions with heterogeneous student populations and interaction patterns, while privacy constraints restrict data sharing and hinder collaborative model development. To address these challenges, this study proposes a federated, explainable learning analytics framework for modeling academic risk trajectories under controlled institutional heterogeneity. The proposed architecture integrates temporal behavioral representations with socio-academic features within a multitask learning scheme, evaluated under both centralized and federated regimes, while modeling institutional heterogeneity through parameterized non-IID partitions that introduce controlled class imbalance, temporal drift, and feature-level variability. Experimental results show that federated models preserve strong discriminative performance and stable convergence as heterogeneity increases. At the same time, calibration metrics exhibit greater sensitivity to distributional shifts, revealing a decoupling between ranking performance and probabilistic reliability. In parallel, explainability analysis shows that the relative importance of features remains structurally stable across institutions, despite variations in contribution magnitudes. Cross-platform evaluation further shows that models retain discriminative capacity when transferred across educational environments, while exhibiting changes in calibration and explanatory intensity. These findings highlight the importance of multidimensional evaluation in federated learning systems, jointly considering performance, calibration, and interpretability, and provide a methodological framework for deploying robust and privacy-preserving learning analytics models in heterogeneous educational settings.