Knowledge Structure-Aware Graph-Attention Networks for Knowledge Tracing

Abstract

Knowledge Tracing (KT) aims to assess learners’ learning states and predict their performance based on prior interactions. However, most existing KT models depend on knowledge concepts instead of specific exercises, leading to the fine-grained information at the exercise level has been ignored, which may weaken the prediction performance of the models. We herein present Knowledge Structure-aware Graph-Attention Networks (KSGAN) for predicting learners’ performance, which uses improved Graph Attention Networks (GATs) to acquire effective exercise representations by taking full advantage of the knowledge structure between knowledge concepts and exercises. Additionally, a representation optimization is devised and integrated into the loss function to alleviate the sparsity of educational data and further improve the prediction performance. Finally, empirical validations on three open benchmark datasets show that our model well outperforms some state-of-the-art models in recent years. Remarkably, our model demonstrates superior prediction performance at exercise level compared to these previous models, without the additional information (e.g., exercise content, temporal information).