Improving Question Recommendation Through Oracle Recommendation Imitation

Question recommendation aims to sequentially recommend personalized questions to improve student learning effects. Most existing question recommenders are built on online reinforcement learning pipelines, which simultaneously collect training data and update the model. While effective for exploration, the collected data is often low quality, limiting the recommender’s ability to learn strategies that truly enhance learning. In this paper, we conduct an empirical study of this issue and propose a new pipeline named Oracle Recommendation Imitation (ORI). ORI first trains a preliminary recommender to generate high-quality samples and uses an oracle selection mechanism to efficiently extract reliable offline data enriched with collaborative information. This oracle dataset, combined with real-time online data, guides the recommender toward effective strategies. Extensive experiments show that ORI, despite its simplicity, can boost the performance of various recommenders by up to several tens of percentage points. Our code is available at https://github.com/wuming29/ORI .