Mixture of Experts-Based Fusion of Aligned Structure and Context for Programming Knowledge Tracing

Programming knowledge tracing (PKT) is an essential yet challenging task that aims to assess students' proficiency in programming skills by tracking their historical performance on programming assignment tasks. Previous approaches typically focus on structural or contextual views of the code to derive corresponding representations to model PKT. Unfortunately, the single-view based code representation may fail to efficiently capture the subtle differences in studentsubmitted code, leading to less effective performance. In this paper, we emphasize the importance of both structural or contextual views of the student-submitted code and formulate PKT as a multi-view fusion task. Guided by two common principles in multi-view learning, namely complementarity and consensus, we propose an efficient Mixture of Experts-based Fusion of Aligned Structure and Context (MoE-FASC) method for PKT. First, to adhere to the principle of complementarity, our MoE-FASC presents an efficient MoE-based structure and context fusion scheme that seeks to learn diverse interactions between structural and contextual representations to fully exploit their complementary information through individual experts collaborating in parallel. Simultaneously, a gating mechanism is employed to dynamically assign appropriate weights for each expert, facilitating an ‘optimal’ aggregation of structural and contextual representations. Second, in pursuit of consensus, we present a knowledge hidden state alignment strategy to ensure a consensus representation between each view. This alignment strategy would capture essential common patterns across different views, facilitating the learning of coherent representations that are unified and less sensitive to view-specific noise, thus enhancing the generalization capacity of our MoE-FASC. We conduct extensive experiments to verify the effectiveness of our MoE-FASC, and the results demonstrate that our proposal achieves better performance compared to the currently popular methods.