Time-aware Session Modeling for Knowledge Tracing

Knowledge Tracing (KT) aims to model learners’ evolving knowledge states from their historical learning records. Most existing KT methods often treat these records as continuous and uniformly distributed. However, we propose that they can be divided into shorter sessions based on temporal distribution characteristics. To address this, we introduce a novel KT model called Time-aware Session Modeling for Knowledge Tracing (TSMKT), which captures knowledge state changes with finer granularity. In particular, we first divide historical learning records into shorter sessions based on temporal distribution characteristics. Subsequently, we present a fine-grained knowledge state modeling component to figure out intra-session and inter-session interaction dependencies and knowledge state changes. Additionally, a global knowledge state modeling component is introduced to holistically model learners’ knowledge states. Experimental results validate the effectiveness of TSMKT.