User Behavior Analysis And Recommendation Using User Community Detection

With the rapid development of Internet technology, various type of web services and applications gradually into people’s lives. Our social activities and information access methods in the past also have changed from the reality society to the virtual network space. In particular, with the rapid expansion of cloud computing, social networks and big data technology, the Internet is full of information and users have become more difficult to meet their needs. As a kind of information service system, recommender system usually needs to construct user model and content model, which is used to describe user interest and information content. Then according to the user’s interests, the content of the information is personalized. The current recommender system has been widely used in music, books, television, tourism and other kinds of live and entertainment services, also has become a very promising develop direction of information technology.Although the past research on the recommendation algorithms have got considerable achievements, but with the actual recommender systems for the long-term operation and the continuous growth of the size of the users and information, the problem of the recommender system has been more serious, and many new problems are exposed. In order to better adapt to the trend of Internet development. In this paper, the algorithm we propose is based on user community detection. And with the help of user network analysis, we optimized the traditional collaborative filtering algorithm, and aim to solve the issues of sparsity of data, interests drift etc. Our research and discussion in this paper are the following aspects below:(1) The construction of user characteristics network based on complex network and the method of user community partition are described in detail. The method divide the large number of sparse user behaviors data into small granularity. The data size is reduced and the speed is promoted.(2) we combine the recommend results of two types of collaborative filtering algorithms, in order to avoid the defects of a certain kind of algorithm, even affect the performance of the whole system. For the Top-N recommendationproblem, the rank evaluation method of recommended items is also promoted. Based on the result of experiments, it shows that the recall and coverage etc. of the proposed algorithm can be significantly higher than the traditional algorithm.(3) Aiming at the problem of users’ interest drift, the rank of user behaviors decay mechanism which rely on time factor is proposed. Due to this mechanism, users’ long-term and short-term behaviors have different influence to their interests, that is also enhance the awareness of user interest in this system. A kind of item popularity attenuation method is adopted in this paper, which can provide more innovative information and improve the system’s diversity.(4) We also adopt a user behavior feedback model based on classification model to filter recommended results. This model treats the users’ behavior of recommended items as latent feedback, divide them into positive and negative samples to train an interest model. Then the system can use this model can be a kind of selector to filter or pick up items in the recommended list. In our experiments, the accuracy of system can be improved when we compared with an algorithm without this model.According to the research above, we also illustrate the implementation of a personalized music recommender system in detail. Include the architecture, technological process, user behavior analysis and so on. This system can provide and push personalized and interesting songs to users based on their history tracks and follow their interests drift due to their feedback.