Matrix Factorization Using LightFM for a Music Recommendation System Based on Emotional and Listening Behavior Awareness

Abstract

This paper summarizes an emotion-aware hybrid music recommendation approach that combines users’ emotional listening characteristics and implicit behavioral signals to enhance personalization under sparse feedback conditions. We construct enriched user profiles by aggregating audio-derived emotional features such as valence and danceability, modeling genre preferences through one-hot encoding, and capturing engagement behavior via skip-rate statistics, followed by systematic preprocessing including outlier removal and normalization. Using standardized emotional features, we apply K-means clustering to assign interpretable mood contexts (e.g., happy, energetic, calm, and sad), which are then incorporated as user-aware signals in a hybrid LightFM matrix factorization model optimized with the WARP-kos loss for ranking-based recommendation. Experimental evaluation demonstrates that the proposed model achieves a Precision@10 of 0.6209, indicating that more than six out of ten recommended tracks are relevant, and a Recall@10 of 0.4663, meaning that approximately 47% of all relevant items are successfully retrieved within the top-10 recommendations. These results highlight the model’s ability to balance accuracy and coverage while outperforming traditional collaborative and content-based baselines, thereby confirming that integrating emotional context with behavioral data significantly improves the effectiveness of personalized music recommendation systems.