Research on Interpretable Knowledge Tracing Model Based on Multi-Feature and Forgetting Behaviors

The objective of knowledge tracing lies in estimating whether a learner will succeed in answering the upcoming question correctly, which is realized by analyzing the chronological log of learner-exercise interactions. As a foundational technique underpinning adaptive learning platforms, its accuracy directly affects the quality of personalized educational services. Nevertheless, a common shortcoming of mainstream approaches lies in their inadequate handling of memory decay, a cognitive phenomenon that progressively erodes mastery as time elapses, thereby producing predictions that diverge from learners' actual performance. To mitigate this shortcoming and lift the predictive ceiling of deep knowledge tracing frameworks, this study proposes an interpretable knowledge tracing approach termed IKT-MFFB (Interpretable Knowledge Tracing with Multi-Feature and Forgetting Behavior), which jointly leverages diverse behavioral signals and a forgetting-aware mechanism. The proposed framework is built upon four sequential stages. In the first stage, behavioral attributes such as answering frequency are mapped into vector form via one-hot embedding, after which the learner-exercise interactions are aggregated into an interaction embedding layer. In the second stage, a composite forgetting factor is engineered: it couples the natural attenuation triggered by elapsed intervals with the consolidation effect arising from repeated practice, thereby aligning more faithfully with empirical findings in cognitive psychology. Comparative experiments performed on three publicly accessible educational benchmarks demonstrate that IKT-MFFB attains markedly superior predictive accuracy over competing baselines, offering a reliable foundation for downstream applications such as personalized recommendation of learning paths.