The Research On Key Technologies Of Recommender System Based On Social Media

With the development of Web2.0and social media, especially the last five years, the popularity of Facebook, Blogger and Twitter, and other network applications, not only created the era of the "National Correspondent", but also brought the information in the field of social media flooding. Faced with the nearly-disaster data, a natural question is:how users can find useful information? Personalized recommendation technology, as an effective solution to "information overload", no doubt, becomes the preferred. A large number of users in social media, the explosive growth of information and data, and the complicacy of user structure, make the recommender system had to face some new challenges, especially the extremely sparse data, real-time recommendation and trusted recommendation, which are three key issues in social media recommended.Around how to overcome the data sparseness problem, speed up the speed of recommendation, improve the reliability of the recommendation and to ensure the accuracy of the recommendation, number of issues involved in the recommender system in the social media environment, are explored and researched in this dissertation. The main research and innovation are as follows:(1) Based on the statistical analysis of user behavior data, collaborative filtering algorithm based on user behavior is presented. Through the analysis of user behavior data, this method can find the global structure and hidden features of the user behaviors, and combine these information and user behavior data as the basis for collaborative filtering recommendation method. The experiments show that the method, to some extent, improves the accuracy of the recommended.(2) For the scarcity data of the user behaviors, semantic-based matrix decomposition prediction methods are presented. The methods extract some semantic information of user behaviors, such as the implicit characteristics of information, context and time information, location information, and use the matrix factorization method to fill in missing data in the matrix of the whole user behavior. Then, the recommended prediction is based on the complement matrix information of user behaviors.(3) For the real-time recommendation, the recommended method based on the clustering of Co-clustering is proposed. First, the method uses Co-clustering method for offline clustering of users and behaviors; And then realizes online real-time recommendation based on the results of the off-line clustering, combining with the recent behavior of the user’s; Finally, the incremental update model continuously updates the user behavior data to ensure the accuracy of the offline clustering results. The method, on the one hand, is to reduce the computational complexity by clustering to reduce the search space of nearest neighbor users; On the other hand, is to reduce the computation time of online real-time recommendation by separating online recommended and the off-line clustering.(4) By using the social relationships in social media, the method is proposed to improve the credibility of recommended. This method combines the real-life social relationships, personal credibility factors in the social network and trust factors between the users, to supplement the similarity model in the original recommendation system, and to achieve the user’s trusted recommendation. The experiments show that the method, on the one hand, can ensure better recommendation accuracy, on the other hand, can effectively prevent the interference of the external factors to a certain extent.