A Learner Ability Profiling Model Based on Dynamic Knowledge Graph and Graph Neural Network

Abstract

To address the insufficient exploitation of structural information, the inadequate modelling of temporal dynamics, and the limited interpretability of ability dimensions in learner ability profiling under open education, this paper proposes a learner ability profiling model based on a dynamic knowledge graph and a graph neural network, termed DKG-LPA. The model first constructs a dynamic heterogeneous knowledge graph comprising student, course and ability nodes, and employs a gated recurrent unit to characterise the cross-period evolution of student states. On this basis, a spatio-temporal graph attention network fuses structural information with temporal information, and a graph-aware ability context mechanism is designed to inject course-ability structural relations into the generation of ability profiles. A memory-augmented knowledge tracing module is further introduced to simulate the accumulation and forgetting of knowledge mastery states. On the Open University Learning Analytics Dataset, the model achieves an AUC of 0.9825, an F1 score of 0.9382 and a mean absolute error of 0.0215, significantly outperforming the baseline methods LightGBM, DKT, DKVMN, AKT and GKT. Ablation studies show that the dynamic graph evolution module contributes most to the overall performance, and that the graph neural network reduces the ability profiling error by 55.6 percent through the ability context mechanism, which validates the effectiveness of each design.