A dynamic multiobjective recommendation method based on soft actor-critic with discrete actions

Recommender systems based on deep reinforcement learning (DRL) have emerged as a research trend due to their ability to capture users’ dynamic preferences and maximize cumulative rewards. Despite their strengths, many DRL-based methods struggle to explore large-scale discrete action spaces effectively, limiting their direct applicability in recommendation tasks. Moreover, recommender systems that focus solely on a single goal (e.g., accuracy) fail to meet the diverse recommendation needs of users. To address these challenges, we propose a dynamic multiobjective recommendation method based on soft actor-critic with discrete actions (DMoR-SACD). First, we utilize gated recurrent units (GRUs) to extract user features and equivalently decompose state inputs into item sequences and interest codes to obtain accurate reward signals. Second, we design a composite reward function including accuracy, satisfaction, and diversity to optimize multi-domain recommendation objectives. Finally, we utilize the reward function to guide SACD to solve the exploration problem in recommender systems. We conduct experiments comparing DMoR-SACD with seven baseline methods. Experimental results show that DMoR-SACD outperforms all baseline methods in terms of cumulative reward, hit rate, normalized discounted cumulative gain, and recommended sequence diversity, suggesting that DMoR-SACD enhances the exploration capability in large-scale discrete action spaces and provides more comprehensive personalized recommendations.