Tag-Based Personalized Music Recommendation

• Main Authors: Wang, Mengsha, Xiao, Yingyuan, Zheng, Wenguang, Jiao, Xu, Hsu, Ching-Hsien
• Format: Conference Proceeding
• Language: English
• Subjects: Collaboration; Collaborative Filtering; Feature extraction; Frequency estimation; Music; Music Recommendation; Recommender systems; Social Tags Mining; Target tracking
• ISSN: 2375-527X
• DOI: 10.1109/I-SPAN.2018.00040

The increasing popularity of internet-accessing smart devices and fast growth of music streaming platforms has greatly changed the way people consume and share music. Thus, music recommender systems got in the focus of research in academia as well as industry. However, the standardization of the number of users listening to music has been directly used in decomposition of collaborative filtering recommendation algorithm, which ignores the user listening to recorded tracks of the actual distribution and leads to failure to fully exploit this information for making the personalization recommendation. In addition, while music content analysis has been an active research topic for decades, the technologies are still in their infancy. The melody, rhythm, timbre and other important characteristics of the music is difficult to extract and process, as a result, the content features of music itself is not fully considered in most of the personalized music recommendation system for users, causing they are not satisfied for the music recommendation. In order to solve the above problems, our method RTCF (Tag-driven Collaborative Filtering Recommendation System), by analyzing the historical records of tracks of users in the data sets, we set up a reasonable scoring mechanism for users based on the training statistical language model (Good-Turing Estimate). In addition, we found that the tag has a related count when describing a song or an artist, which means that different tags can describe the characteristics of the track and the artist in varying degree. In combination with our analysis of the listening records of each user, the tags not only serve as additional information to broaden the content of the music but build the personalized music recommendation for users. According to the experimental results of users on the dataset containing tags, it indicates that our proposal method has a great advantage over other tag-based methods in making personalized recommended performance.