DGKT: Denoising Student Knowledge states via Generative Diffusion for Robust Knowledge Tracing
Xinyu Nie, Yu Su · 2025
Knowledge Tracing (KT) aims to model students’ evolving knowledge states based on their learning interactions. While deep learning-based KT models have achieved strong predictive performance, they are often vulnerable to noise in educational data, such as guessing, slipping, or low-quality exercises, which can severely distort the estimation of student knowledge. In this paper, we propose DGKT, a novel denoising generative knowledge tracing model inspired by diffusion probabilistic models. Specifically, we construct a concept-aware memory module to capture interactions between questions and skills and estimate their influence on learning. Then, we introduce a generative denoising mechanism to transform noisy latent states into cleaner knowledge states. The final denoised representation is used to predict future student responses. Experiments on three real-world KT benchmarks demonstrate that DGKT consistently outperforms best-performing baselines, especially under noisy conditions, highlighting its robustness and generalization capability.