Improving Music Recommendation System by Applying Latent Topics of Lyrics

Abstract

The proposed music recommendation system was developed by using various information filtering approaches based on user context and song context. This study proposes a music recommendation system with Latent Dirichlet Allocation (LDA) by using user listening behavior and analyzing a latent relationship of each song. As a consequence, small musical niche genres without listing history will become a member of their respective topic groups. Modeling topic analysis of LDA is utilized for songs lyric as well as the user action and, then song group preferences support the collaborative filtering recommendation engine. The system addresses the optimization of the cold start problem of adding new items in Collaborative Filtering by lyric analysis with LDA. Predicted ratings for user recalculated by combination matrix of song listening action with binary rating values ​​and latent topic group result of lyrics. In this analysis, a system proposition compared with two models, normal collaborative filtering and user defined genre group preference.