Abstract
Abstract
Knowledge tracing, which models the evolution of student knowledge from interaction sequences, is fundamental to adaptive learning systems, yet existing approaches struggle to achieve scalability, computational efficiency, and interpretability simultaneously. This paper presents EKT-XAI, an integrated framework addressing these three challenges jointly through: (1) a distributed preprocessing pipeline that processes 95.3 million EdNet-KT1 interactions in approximately 7 minutes; (2) a lightweight transformer architecture of 1,195,809 trainable parameters (4.56 megabytes) that performs inference in 8.4 milliseconds per sequence on a single processor core, without hardware acceleration; and (3) a multi-level interpretability module combining attention visualization, skill difficulty analysis, and learning trajectory tracking, available at prediction time without additional inference passes. On EdNet-KT1, the framework attains an area under the ROC curve of 0.6881, performing comparably to the strongest baseline, Deep Knowledge Tracing (0.6917), and above four further deep learning models and all traditional machine learning baselines. Deep Knowledge Tracing is the more compact model but offers no interpretability mechanism. An independent replication on a freshly drawn sample yields 0.6890, indicating that differences below approximately 0.005 on this configuration should not be over-interpreted. Cross-dataset evaluation on the Open University Learning Analytics Dataset for dropout prediction yields 0.8475, competitive with tree ensembles (random forest 0.8436, XGBoost 0.8412) on a task whose structure differs markedly from sequential knowledge tracing. The novelty of this work is architectural in the systems sense rather than the algorithmic one; no new attention mechanism or explainability technique is proposed. Explanation faithfulness and pedagogical utility are not empirically validated and are stated as limitations.