出版社：The Japanese Society for Artificial Intelligence
摘要：This paper proposes a method for recommending music items without explicit feedback. Context and content features are used as auxiliary information to compensate for implicit feedback. The recent development of communication technology and portable electronic devices has changed the way of consuming music. We can access a vast amount of music items via online music streaming services. As a result, finding appropriate music items from enormous resources gets to be difficult for users. To help users discover their favorite music items, recommender systems in the music domain have been studied. This paper focuses on two challenges specific to music recommender systems: the difficulty of obtaining explicit feedback such as ratings, and the importance of making use of context information. To handle the context information as auxiliary information to compensate for implicit feedback, this paper employs FMs (Factorization Machines), in which the context information is treated as features. Utilizing the merit of FMs that can easily introduce features, this paper also introduces content features in addition to context features. As it is known that using low-level content features directly is not effective because of the semantic gap, this paper proposes two types of abstract content features: UGP (user genre profile) and UCP (user content profile). The effectiveness of the proposed method and the effect of negative sampling methods are evaluated in terms of MPR with #nowplaying-rs and LFM-1b dataset. The result of the experiment shows that the proposed method outperforms wALS (weighted Alternating Least Squares), which is one of the popular music recommendation algorithms based on matrix factorization. The characteristics of the proposed sampling methods are investigated with different settings of the parameters and the ratio of negative samples. As for the effectiveness of each feature, it is found that a feature is effective when JS (Jensen-Shannon) divergence of popularity distribution among different feature values is large. It is also shown that the UCP and UGP cluster labels are more effective than using content features directly.
