Balancing Accuracy and Beyond-Accuracy Indicators in Course Recommendation through Multi-Objective Reinforcement Learning

It is crucial to balance accuracy and beyond-accuracy indicators in recommendations through multi-objective optimization (MOO) methods. Existing multi-objective recommendation methods have several issues: they cannot dynamically model users' long-term and short-term preferences; they fail to solve non-convex problems and are prone to getting stuck in local optima; they neglect the issues of diversity and novelty in recommendations, etc. Therefore, this paper proposes using multi-objective reinforcement learning (MORL) to address the trade-off between accuracy, diversity, and novelty in course recommendation. First, dynamic modeling of users' long-term and short-term interest preferences is conducted using user feature information, course feature information, and users' historical course features. Second, the Actor-Critic network is applied to course recommendation, and a multi-layer reward function and Q-value function are introduced to learn multiple objectives simultaneously. Diversity and novelty assessment indicators are also added to measure the effectiveness of course recommendations Finally, experiments are conducted on three real MOOC datasets, proving the superior performance of the proposed course recommendation model in dynamically modeling user preferences and balancing accuracy and non-accuracy aspects.