Presentation Information
[1Yin-A-65]Research on Emotion-Transition Music Recommendation System Based on the ISO Principle and LLM-Based Audio Feature EstimationAiming for Active Transformation of Psychological States via Music
〇Yinuo Fan1 (1. Yokohama International School)
Keywords:
Affective Computing,Music Recommendation Systems,Learning Support,Mental Healthcare
This study proposes MindTune, a mobile music recommendation system that actively guides users from their current emotional state to a desired target state using curated playlists. Unlike Spotify’s history-based recommendations, MindTune applies music therapy’s ISO principle and Russell’s circumplex model to design emotionally adaptive transition paths. The system implements four strategies (Arousal First, Valence First, Linear, and Dynamic) using a scoring algorithm for 1,368 tracks. Addressing the 2024 deprecation of Spotify’s Audio Features API, this research introduces an LLM-based approach to estimate audio features from track metadata. Validation against Spotify’s original data showed strong correlations for energy (r=0.84) and acousticness (r=0.81), while the happiness feature (used as a proxy for Spotify’s valence) showed moderate correlation (r=0.41), indicating room for improvement in the valence dimension. Comparative experiments demonstrated that the Dynamic strategy (mean rating 4.3/5) outperformed the Linear strategy (3.3/5) in emotional regulation, achieving smoother transitions and higher user empathy. These results suggest that combining LLM-based audio feature estimation with emotion-transition algorithms is a promising direction for personalized music recommendation systems.