Reproduce “A Bandit-Based Approach to Educational Recommender Systems: Contextual Thompson Sampling for Learner Skill Gain Optimization” on OpenResearch

About this paper

Lukas De Kerpel, Arthur Thuy, Dries F. Benoit

A contextual bandit approach, Linear Thompson Sampling (LinTS), was developed for educational recommender systems to optimize learner skill gain by personalizing exercise sequences. Evaluated on the ASSISTments 2017 dataset, LinTS achieved a 20.7% improvement in average skill gain over User-Based Collaborative Filtering, demonstrating superior performance by effectively balancing exploration and exploitation.

Key insights

• Contextual bandit algorithms, particularly LinTS, offer a more adaptive and principled approach to personalized learning than traditional CF methods, effectively balancing exploration of new exercises and exploitation of known effective ones.
• Optimizing directly for pedagogically relevant metrics like 'skill gain' (derived from BKT) ensures the recommender system's objective is aligned with genuine knowledge acquisition and competence development.
• LinTS exhibits an adaptive exploration-exploitation dynamic, initially sampling broadly to gather information and then shifting towards focused exploitation of high-value learning opportunities as evidence accumulates.

Summary by alphaXiv.