Triple Masked Attention Network with Skill Embedding for Knowledge Tracing

• Chen Wang,
• Bo Xie,
• Peiqian Zhan,
• Xiaoqi Yang,
• Hua Zhang,

Vol. 19, No. 3, March 31, 2025
10.3837/tiis.2025.03.004, Download Paper (Free):

• knowledge tracing
• skill tags
• masked cross attention
• Transformer block
• response prediction

Abstract

Knowledge Tracing aims to uncover learners' mastery levels of different knowledge or concepts based on their interaction sequences. It has attracted wide applications in intelligent educational systems. Recent methods based on deep neural networks have shown promising performance in this task due to their powerful feature learning capabilities. However, these approaches often overlook skill tag information, making them insufficient for personalized learning. To address this issue, we propose the Triple Masked Attention Network with Skill Embedding for Knowledge Tracing (TMAN-SEKT), which considers the embedding of skill tags in addition to question-response sequences. We employ a Transformer block with rectified masked self-attention to individually capture contextual representations of three input sequences, including questions, question-responses and skill tags. Then, another Transformer block with triple masked cross attention is presented to deeply explore the contextual relationships among these three input sequences. We further design a variant of cross-attention to address the sequence heterogeneity when computing attention scores. Extensive experiments on the EdNet benchmark dataset demonstrate that our TMAN-SEKT consistently outperforms previous state-of-the-art knowledge tracing methods.

Statistics

Cite this article