MTKT: Learning Multi-granularity Temporal Characteristics for Attention Based Knowledge Tracing
We added MTKT into our pyKT package.
The link is here and the API is here.
Original paper can be found at Bai, Youheng et al. “Learning multi-granularity temporal characteristics for attention based knowledge tracing.” Neurocomputing, 2025.
Title: Learning multi-granularity temporal characteristics for attention based knowledge tracing
Abstract: Knowledge tracing (KT) plays a crucial role in personalized learning by predicting students’ future performance based on their historical interaction data. Despite advancements in KT models, they often struggle to effectively model the intricate temporal dynamics inherent in student learning processes, particularly in capturing multi-granular temporal information, accommodating irregular time intervals, and handling varying sequence lengths within student interaction data. To address these challenges, we propose MTKT to model multi-scale time patterns and long-term behaviors in student interaction data. Specifically, MTKT introduces three key components. First, a multi-aspect embedding layer captures both temporal information and interaction behaviors from student learning data. Second, a dual attention module with linear decaying bias applies negative bias to attention scores, effectively modeling long-term forgetting in students’ knowledge states across varying sequence lengths. Finally, a causal interaction convolution module captures fine-grained representations in both short-term and long-term interactions. We evaluate MTKT on three public educational datasets and compare it with state-of-the-art KT models. Our experiments show that MTKT consistently outperforms existing models in student performance prediction and length-adaptive prediction.
