EXPLORE -- Explainable Song Recommendation

TL;DR

The paper tackles opaque song recommendations and limited user control by proposing an explainable, mood-aware music recommender. It adopts a hybrid of collaborative filtering and content-based methods using Explainable Matrix Factorization, mood-driven UI, and Spotify/MLHD data. Evaluation via $RMSE$ and user studies, along with ranking metrics $MAP@K$ and $NDCG@K$, shows promising predictive accuracy and user satisfaction. Key contributions include interpretable latent-factor mappings to song attributes, interactive mood controls, and out-of-corpus recommendations, with pathways identified for scalability and group recommendations. Overall, the work advances transparent, user-centric music recommendations with practical implications for deployment in streaming contexts.

Abstract

This study explores the development of an explainable music recommendation system with enhanced user control. Leveraging a hybrid of collaborative filtering and content-based filtering, we address the challenges of opaque recommendation logic and lack of user influence on results. We present a novel approach combining advanced algorithms and an interactive user interface. Our methodology integrates Spotify data with user preference analytics to tailor music suggestions. Evaluation through RMSE and user studies underscores the efficacy and user satisfaction with our system. The paper concludes with potential directions for future enhancements in group recommendations and dynamic feedback integration.

Figures (7)

• Figure 1: Data Integration Flow Chart
• Figure 2: Product User Interface
• Figure 3: Mood Control Bar
• Figure 4: User Neighbour Interaction View
• Figure 5: RMSE-MSE Evaluation