Reproduce “MUSE: Music Recommender System with Shuffle Play Recommendation Enhancement” on OpenResearch

About this paper

Yunhak Oh, Sukwon Yun, Dongmin Hyun, Sein Kim, Chanyoung Park

KAIST researchers developed MUSE, a music recommender system that uses self-supervised learning with transition-based augmentation to improve recommendations, particularly in shuffle play sessions. The system consistently enhanced recommendation accuracy across Spotify's Music Streaming Sessions Dataset, showing a 3.00% Recall@10 improvement over its backbone model in shuffle play.

Key insights

• Shuffle play sessions, often considered problematic due to their randomness, can be effectively leveraged for recommendation if unique transition patterns are specifically addressed through tailored augmentation.
• Domain-specific augmentation, like transition-based insertion based on global track transition frequencies, is crucial for mitigating the impact of high entropy in sequential data.
• A self-supervised learning approach with both item-level and similarity-based matching strategies enhances the robustness and representational power of recommender systems across different session types.

Summary by alphaXiv.