Contrastive Learning Based on Bipartite Graphs for Interpretable Knowledge Tracing

Kai Chen, Shun Mao, Qiwen Zheng, Yuncheng Jiang

Abstract: The use of online educational platforms has become an increasingly popular way for people to acquire knowledge from the Internet. Knowledge tracking tasks are defined as predicting learners’ future performance based on their historical interaction data. Many researchers employ the bipartite graph representation of the relationship between two different types of nodes, exercises and knowledge concepts, to make better predictions. However, the implicit associations between nodes of the same type in this graph cannot be well captured, which limits the model’s ability to extract deep relationships and semantic information, resulting in poor generalisation ability. To address this limitation, we propose a bipartite graph-based contrastive learning method with dual-view data augmentation, which captures implicit associations between nodes of the same type while improving the representation of relationships between exercises and concepts. In addition, the application of deep learning methods in knowledge tracing often suffers from poor interpretability of prediction results. To address this issue, this paper incorporates the mastery scores of exercises and knowledge concepts into the predictions to improve the interpretability of the model. Experimental results show that our model outperforms related models on several datasets.