A lightweight knowledge tracing method for multi-source heterogeneous data based on LLMs and GANs

Knowledge tracing (KT) is essential for personalized education but it faces an unresolved tension between the need to model complex, multi-source heterogeneous data and the requirement for real-time computational efficiency. This paper proposes LG-MKT, a lightweight KT method integrating Large Language Models (LLMs) and Generative Adversarial Networks (GANs). LG-MKT utilizes LLMs offline to construct semantic anchors, ensuring cross-source consistency without inference overhead. Simultaneously, a GAN-based regularizer harmonizes latent distributions to mitigate data sparsity. Experiments on ASSIST09, EdNet, and MOOCCube show LG-MKT outperforms baselines with 0 . 31 ms inference latency. Specifically, the model achieves an AUC of 0.845 on MOOCCube and reduces error metrics like MAE by up to 8.76%, demonstrating a superior balance between predictive accuracy and real-world deployability.