Deep Knowledge Tracing for Explainable Problem Recommendations on Codeforces

James Zhao, Fang Sun, Yizhou Sun

Keywords: prerequisite constraints, causal recommendation, knowledge tracing, deep knowledge tracing, explainable recommendation, educational recommender systems, transformer models

Abstract: Contemporary competitive programming platforms, such as Codeforces, offer a vast array of problems, which can overwhelm novice users encountering competitive programming for the first time. This complexity highlights the need for intelligent recommendation learning paths. However, designing these paths using traditional recommendation systems poses significant challenges, as many do not incorporate temporal knowledge and prerequisite concepts, often relying solely on correlation-based methods. We approach this issue from the perspective of deep knowledge tracing (DKT), utilizing transformers for predicting users' skill levels and recommending problems. Our model employs DKT to learn a dynamic knowledge vector that predicts the probability of users successfully solving any given problem. Additionally, we enhance our DKT transformer architecture with a TransE-based prerequisite graph. Our model achieves an ROC AUC score of 0.75 for knowledge tracing, paving the way for explainable recommendations. Users who interact with our system benefit from real-time insights into their weaknesses while receiving targeted suggestions to improve their knowledge.

Submission Number: 15