Mutil-View Contrastive Learning for Knowledge Tracing Based on Multi-Faceted Exercise Feature Fusion

To address the issues of inadequate consideration of multifaceted features of exercises and dependence on large scale labeled data in existing deep knowledge tracing models, this paper proposes a multi-view contrastive learning method for knowledge tracing based on multi-faceted exercise feature fusion. This method achieves the extraction and integration of internal and external features from various aspects of exercises by constructing four views: the relationship between the exercise knowledge concept, the correlation of the exercise, the difficulty of the exercise and the discrimination of the exercise. By incorporating positive and negative instances from four views along with contrastive learning, this method enhances the relevance of similar exercise representations at both content and evaluation levels, thus obtaining high-quality latent state representations of student knowledge mastery. Experimental results demonstrate that this approach significantly improves model performance and predictive accuracy, particularly in limited data scenarios with education benchmark data set.