Dynamic comprehensive difficulty knowledge cells based on KAN network and stable learning for knowledge tracing

Knowledge Tracing (KT) is a deep learning task aims at tracing students' knowledge states and predicting their future performance. However, existing methods overlook the impact of dynamically changing comprehensive difficulty on students' knowledge states, leading to suboptimal performance in real-world educational scenarios. Additionally, students' correct responses may sometimes result from lucky guesses, making it difficult for models to accurately update comprehensive difficulty and effectively trace students' knowledge states. To address these challenges, we propose DDKC, the novel Dynamic comprehensive Difficulty Knowledge Cells, which incorporates two key components: (1) the dynamic comprehensive difficulty knowledge cells incorporating KAN network, to dynamically update the comprehensive difficulty according to students' evolving knowledge mastery level; (2) a novel training strategy based on Stable Learning, to mitigate spurious correlations caused by students' lucky guesses, and focus more on tracing their real knowledge states. We evaluate DDKC on several public educational datasets, demonstrating that it surpasses current state-of-the-art models, and exhibits excellent interpretability in capturing dynamic changes in comprehensive difficulty and eliminating spurious correlations.