Multi-level contrastive graph learning for academic abnormality prediction

Academic Abnormality Prediction aims to predict whether students have academic abnormalities through their historical academic scores. However, existing research methods still have the following challenges: (1) Student behavior. Only the students' historical academic performance is considered, ignoring the impact of student behavior in student status. (2) Data imbalance. The number of academically abnormal students is much less than that of ordinary students, resulting in a data imbalance problem. Therefore, in response to the above challenges, this paper proposes a Multi-level Contrastive Graph learning for academic abnormality prediction (MCG). Specifically, firstly, we capture student behavior and fuse it with student historical achievement data based on a Graph Neural Network (GNN), Thereafter, we construct an embedding space for sample interpolation, which generates virtual nodes of abnormal students, thereby alleviating the data imbalance problem. Moreover, we introduce a multi-level contrastive learning module to precisely learn node representations and maximize the consistency between different views of the same node in the target and online networks for data augmentation. Experiments on real datasets show that the abnormality prediction performance of MCG outperforms the existing state-of-the-art methods.