Adaptive course recommendation in MOOCs

Abstract In the process of course learning, users incline to change their interests with the improvements of their cognition. Existing course recommendation methods usually assume that users’ preferences are static. They fail to capture the user’s dynamic interests in sequential learning behaviors. In this respect, the recommendations show low accuracy and adaptivity, especially when users have diverse interests in many different courses. Thus, they may not be suitable for applying in the online course recommendation scenario. In this paper, we propose a novel course recommendation framework, named Dynamic Attention and hierarchical Reinforcement Learning (DARL), to improve the adaptivity of the recommendation model. DARL automatically captures the user’s preferences in each interaction between a profile reviser and a recommendation model, and thereby enhances the effectiveness of course recommendation. For tracking the changes in users’ preferences, DARL adaptively updates the attention weight of the corresponding course at different sessions to improve the recommendation accuracy. We perform empirical experiments on two real-world MOOCs (i.e., Massive Open Online Courses) datasets. Experimental results demonstrate that DARL significantly outperforms state-of-the-art course recommendation methods in terms of major evaluation metrics.