講演情報
[4O5-IS-2c-02]Implicit Graph Learning with Adaptive Temporal Bias for Knowledge Tracing
〇Sunil Kumar Maurya1, Shuko Sasaki2, Makoto Kawano1, Yusuke Iwasawa1, Yutaka Matsuo1 (1. University of Tokyo, 2. Matsuo Institute)
regular
キーワード：
Knowledge Tracing
Knowledge tracing seeks to model how a learner's mastery evolves over time based on sequences of question–response interactions. Recent graph-based approaches represent concept dependencies as structured graphs and leverage message passing or attention mechanisms to capture learning dynamics. However, these methods typically depend on predefined or loosely learned graphs and lack a principled way to jointly account for temporal distance and contextual relevance when aggregating past interactions. This may result in amplifying uninformative signals while underweighting meaningful recent context. In this work, we introduce a graph attention-based knowledge tracing model that reframes the problem as implicit graph learning over a causal directed acyclic graph. Our model incorporates a context-aware, learnable temporal bias that modulates the importance of historical interactions according to their distance from the current prediction step. We further incorporate an efficient channel recalibration module that performs lightweight cross-channel reweighting to suppress noisy feature dimensions introduced during interaction aggregation, enhancing prediction robustness. Extensive experiments across six benchmark datasets demonstrate that our model consistently outperforms existing baselines.