Research And Implementation Of A Music Recommender System Based On Multidimensional Time Series Analysis

With the rapid growth of mobile-Internet and smart-phone in recent years, many music radio applications come to our lives, such as Last.fm, Pandora and douban.fm. The recommender algorithms behind most of these applications do not take the rela-tionship between users’ listening behaviors and the context information into consid-eration leading to unsatisfactory recommendations. There exist a few algorithms that take the relationship into consideration; however, these algorithms do not handle the time-related features of users’ behaviors in a reasonable way. We believe that users’ future behaviors are simultaneously affected by their long-term behaviors, mid-term behaviors and immediate behaviors. To make full use of these time-related features, we propose a music recommender method based on multidimensional time series anal-ysis. Our method first exploits users’ mid-term behavior, and then incorporates users’ mid-term behavior with their long-term behavior and immediate behavior. Moreover, we implement a system prototype to preliminarily verify the feasibility of our method. Our main contributions are summarized as follows:1. A music recommender method based on multidimensional time series analysis was proposed. The method first uses topic models to model a song as a proba-bility distribution over some latent topics and then models a user’s behavior as a multidimensional time series. By analyzing the time series, the method can forecast the user’s future behaviors and get reasonable recommendations.2. A music recommender method based on users’ long-term behaviors, mid-term behaviors and immediate behaviors was proposed. The method can dynamically adjust the weights of these behaviors and get better recommendations than the method purely based on multidimensional time series analysis. 3. A music recommender system prototype based on the above methods was imple-mented. In order to improve the efficiency, we use Storm, a distributed realtime computation framework, to implement the system and achieve the desired effect-s.