The Research On Technologies Of Music Recommending Based On Implicit Feedbacks

With the rapid development of Web2.0technology, people can upload and share musiconline freely, which led to an explosive growth of information. The millions of free onlinemusic create challenges to both consumers and music providers. This paper proposes twonovel approaches for music recommendation based on the analysis of implicit feedbacks. Theproposed approaches can solve the cold start problem and produce more accuraterecommendations as well as make the recommender more understandable. This main work ofthis paper includes two parts of content.Firstly, based on the weighted social tags, an enhanced content-based musicrecommending method is proposed. In this method, each listened song was classified into likeor unlike set according to user’s playing behaviors, then the songs’ social tags were collectedfrom LastFm website and weighted according to their frequency in the collected tags, finallythe user’s preference degree for each song was quantified with the weighted tags, and thecandidate songs wit high preference degrees would be recommended to him. Theexperimental results show that the skip hit rate, precision, recall and F1-measure improvesabout6.12%,1.68%,0.748%and0.751%respectively.Secondly, to solve the cold start problem, a hybrid music recommending method basedon social tagging and user attributes is proposed. In addition, an attributes inference methodbased on implicit feedback is proposed in order to solve the attribute information sparsity. Inthis method, establish a TFIDF attribute classification model by analyzing the known attributeusers’ listening records, and then use the model to infer those unknown users’ attributes.Based on the attributes, an attribute filtering method is built in order to start new users.Finally linearly combine the attribute filtering and content filtering to recommend music forusers. The attribute inference experimental results shows that the precision for infer age, sexand country are91%,99%and93%respectively. And the music top-k recommendationexperimental results shows that the precision of the hybrid method achieves98.5%, which hasbeen improved0.46%compared with content filtering method.In conclusion, this paper proposed two methods to provide high accurate andunderstandable music recommendation. With such work, it is beneficial to give users moreaccurate and diverse music service.