Enhancing Personalized Performance through a Deep Reinforcement Learning-Based Recommendation System

Recommender systems have become vital in managing information overload and boosting user engagement. However, traditional methods rely on fixed user preferences inferred from past interactions, often overlooking the dynamic nature of user behavior. This study introduces a fresh perspective by framing the recommendation process as a slate Markov Decision Process (slate-MDP) and leveraging deep reinforcement learning (DRL) to adapt recommendations through ongoing interactions. The ultimate goal is to enhance user engagement over extended periods in the face of evolving environments. The research involved simulating various preference scenarios and evaluating the performance of two DRL-based algorithms: FullSlateQ, an undecomposed full slate Q-learning approach, and SlateQ, a Deep Q-Network (DQN) with slate decomposition approach. The findings highlight SlateQ’s impressive 10.57% superiority in non-static environments compared to FullSlateQ. Furthermore, both algorithms tend to adopt a short-term approach when employing a moderate discount factor, struggling to optimize long-term user engagement. The application of deep reinforcement learning in dynamic settings showcases substantial advancements in user engagement, emphasizing the real-world relevance of our discoveries.