Dual-Channel Knowledge Tracing With Self-Supervised Contrastive and Directed Interaction Learning

With the widespread adoption of the internet, intelligent tutoring systems have experienced rapid advancements in education. Knowledge tracing (KT) aims to predict students’ future performance by modeling their behavioral sequences. While existing models have made notable progress in some areas, challenges remain in fully uncovering the higher-order relationships among exercises, concepts, and students. To address these limitations and effectively capture the intricate dependencies between learning behaviors and exercises, this paper introduces a novel approach: Dual-Channel Knowledge Tracing with Self-Supervised Contrastive and Directed Interaction Learning (SCDI). This method enhances predictive accuracy by leveraging graph-structured embeddings of interaction data, offering a more comprehensive analysis of the learning process. SCDI consists of two distinct channels. The first channel constructs a heterogeneous graph with student, exercise, and concept nodes, using a Meta-path Aggregated Graph Neural Network (MAGNN) to extract high-order information from exercise and concept nodes. It also incorporates a self-supervised graph contrastive learning mechanism to improve the quality of node embeddings. The second channel builds a directed graph from interaction sequences, using Graph Attention Convolution (GATC) to capture dynamic relationships and model how interactions evolve over time. Features from both channels are integrated and passed to a downstream KT model to predict students’ future performance. Experimental evaluations on four datasets demonstrate that SCDI significantly enhances the accuracy of downstream KT models, establishing its effectiveness and robustness.