CausalSTKT: Disentangled Spatiotemporal Heterogeneous Graph Learning for Robust Knowledge Tracing Under Distribution Shifts

Knowledge tracing supports adaptive intelligent tutoring systems by estimating learners’ evolving knowledge states from sequential interactions. However, existing models often rely on dataset-specific correlations and may suffer substantial performance degradation under distribution shifts. This paper proposes CausalSTKT, an SCM-guided knowledge tracing framework that integrates spatiotemporal modeling over a global item–knowledge-component bipartite graph with disentangled representation learning. A relation-aware graph encoder captures high-order dependencies among items and knowledge components, while a temporal transition module models the evolution of learner states. To reduce sensitivity to environmental variation, the learned representation is decomposed into a mastery component and an environment component using a directional decorrelation penalty. An environment-resampling mechanism, which recombines the mastery component of one learner–step pair with the environment component of an unrelated pair drawn at random within the mini-batch, is further introduced to encourage predictions that are stable across environments. We also derive an out-of-distribution risk bound showing that a smaller representation-entanglement residual leads to a tighter generalization bound under the stated assumptions. Experiments on five real-world educational datasets demonstrate that CausalSTKT achieves a mean AUC of 0.8371 and a mean accuracy of 0.8136, exceeding the strongest baseline on each metric by 3.42 and 3.61 percentage points, respectively. In the evaluated zero-shot cross-dataset transfer settings, its average AUC drop is 9.1%, compared with 20.0% and 21.1% for the two baselines evaluated under the same protocol. These results indicate that CausalSTKT provides a robust computational approach to knowledge tracing in intelligent learning systems affected by distribution shifts.