A latent representation of users, sessions, and songs for listening behavior analysis

Understanding user listening behaviors is important to the personalization of music recommendation. In this paper, we present an approach that discovers user behavior from a large-scale, real-world listening record. The proposed approach generates a latent representation of users, listening sessions, and songs, where each of these objects is represented as a point in the multi-dimensional latent space. Since the distance between two points is an indication of the similarity of the two corresponding objects, it becomes extremely simple to evaluate the similarity between songs or the matching of songs with the user preference. By exploiting this feature, we provide a two-dimensional user behavior analysis framework for music recommendation. Exploring the relationships between user preference and the contextual or temporal information in the session data through this framework significantly facilitates personalized music recommendation. We provide experimental results to illustrate the strengths of the proposed approach for user behavior analysis. © Chia-Hao Chung, Jing-Kai Lou, Homer Chen.