Fine-grained causal effect estimation of learning resources via dynamic causal heterogeneous graph neural networks

Abstract

Abstract Effect evaluation of learning resources in online education platforms is crucial for optimizing instructional strategies and enabling personalized learning. However, conventional methods struggle to address the inherent challenges posed by the heterogeneous graph structure of student-resource interactions, dynamic temporal patterns, and confounding bias in causal identification. To tackle these issues, we propose a Dynamic Causal Heterogeneous Graph Neural Network (DCHGNN), an end-to-end causal inference framework. DCHGNN comprehensively models the complex and time-evolving relationships among students, resources, and assessment activities by constructing dynamic heterogeneous graphs. It further integrates advanced graph representation learning with a doubly robust estimator to accurately and robustly estimate the Average Treatment Effect (ATE) of learning resources while mitigating selection bias (a core confounding factor). Extensive experimental results including ablation studies, sensitivity analysis, and comparisons with published algorithms on real-world educational data demonstrate that DCHGNN achieves more accurate and robust causal effect estimation compared to traditional baselines, successfully revealing the differential causal impacts of various resource types. The proposed framework shows significant promise for data-driven educational decision-making, facilitating the effective allocation of learning resources and the enhancement of overall teaching efficacy.