Policy GRU-RL: Simplified Music Playlist Recommendation Using Sequential on Reinforcement Learning Concept

In the realm of streaming services, recommendation systems play a crucial role in meeting user preferences by aiding them in discovering music tailored to their tastes. Reinforcement learning (RL) stands out as a popular method for music recommendations. Nevertheless, prior approaches have grappled with the challenge of over-fitting. After a certain learning period, the agent may struggle to predict actions solely based on past interactions, posing issues for the current user context. To address this limitation, previous methods must be retrained by resetting all parameters in the agent. This study introduces the Policy GRU-RL method, which combines sequential-based learning and reinforcement learning to tackle over-fitting without the necessity of resetting all parameters. This method capitalizes on the features of a recurrent network by implementing an epsilon-greedy policy within the GRU gate. An update gate in the GRU determines whether to choose the random action (current input of the GRU cell) or the optimal action (information from the preceding GRU cell, containing actions with maximum rewards). Additionally, it carries " and action values through each iteration, assessing over-fitting by checking for duplicated predicted actions in specific i terations. Subsequently, the " parameter in the agent is reset. The results demonstrate that our proposed Policy GRU-RL surpasses baseline approaches in terms of accuracy.