Presentation Information
[2Yin-A-37]Promoting Diverse Music Consumption via Simulation-based Reinforcement Learning in Automated Playlist GenerationA Study on Enhancing Music Listening Diversity on Streaming Platforms
〇Takuya Yamauchi1 (1. Office Yamauchi)
Keywords:
reinforcement learning,Music Recommendation System,Double DQN / Dueling DQN,User Modeling,Sequential Decision Making
This study proposes reinforcement learning-based music recommendation systems using several models, including a combination of Double DQN (Double Deep Q-Network) and Dueling DQN architectures. It learns the next song to recommend as an action, using the user's music listening history as the state. Furthermore, it pre-trains using an LSTM-based user simulator that models users listening to songs, stabilizing agent learning by generating realistic user responses. This study constructed an environment capable of handling both real data and deterministic user simulators, enabling playlist generation with duplicate avoidance and diversity bonus features during inference, and evaluated its performance.