Semantic-tailored Variational-Contrastive Graph Learning for Cognitive Diagnosis

Cognitive diagnosis (CD), the foundation of intelligent education, aims to assess students' cognitive levels in knowledge concepts. Graph-based CD enhances diagnostic performance by incorporating high-order relations among entities, such as students, exercises, and knowledge concepts. Recently, self-supervised learning has been applied to CD to address data sparsity. However, existing contrastive learning methods may distort the student-exercise graph and overlook important semantic heterogeneity between correct and incorrect response logs. To address these limitations, we propose the Semantic-tailored Variational-Contrastive Graph Cognitive Diagnosis (SVGCD) method. First, a semantic-aware GNN is used to generate entity representations for different semantic environments. Then, a semantic-specific variational graph reconstruction module infers representation distributions and reconstructs semantic subgraphs while preserving the original graph structure. Additionally, a semantic-specific contrastive strategy introduces high-quality self-supervised signals while retaining semantic characteristics, enhancing student modeling for CD. Extensive experiments on two real-world datasets validate the effectiveness of our SVGCD. The code is available at https://github.com/XChuckie/SVGCD.