Researches Of Recommendation Algorithms Based On User Decomposition And Social Fusion

Being in the era of internet, people are enjoying the information and service provided by the internet without temporal and spatial limitation, meanwhile, they have to face massive and growing scale data what involves almost invalid data. As an important tool in filtering information and providing personalized services, recommendation systems become more and more popular and essential. Typically, recommendation techniques are based on collaborative filtering, which automaticdally predicts interests of an active user by leveraging rating information from similar users or items. Due to their application values and academic challenges, recommendation systems have been extensively studied.However, the mixture of user history behaviors among multiple users, or the interaction of users’interests degrade the recommendation performance. Many existing approaches to recommendation systems can’t accurately capture the user preferences portrait. Hence, this thesis try to represent a user’s preferences by introducing virtual user, and a virtual user is vectors of profiles which are exploited to represent user’s interests and help generate recommendations. To address the two issues, the objective of this thesis is to study how to decompose and fuse multiple users’preferences to improve the recommendation performance, respectively.In this thesis, we first propose two user decomposition algorithms to study the decomposition of the mixed behaviors of multiple users, and make accurate predictions for these users. We adopt recommendations in internet protocol television (IP-TV) service as our target, because this service is indistinctly shared by family members. The first method is named as "User identification for IP-TV recommendation based on periodical events dection". This method defines a virtual user as activities over a time slot. The implicit rating technique is exploited to capture preferences of virtual users, and similar virtual users are extracted as members. Recommendations are provided according to the captured user preferences. The second method is named as "User identification for IP-TV recommendation based on subspace clustering". Due to the performance of first method is greatly limited by given time slots, thereforce, the second method clusters the time subspace based on the factorization of user-item-time tensor. Experimental results show the effectiveness of both methods and the second one is better.In social networks, users typically communicate with each other on kinds of things, this provides an ideal environment to study the fusion of user’s interests. Hence, based on the assumption that a user (or a user’s preferences for items) are affected by social neighbors of this user, two novel social fusion algorithms are further proposed in this thesis. The both algorithms fuse the user-user social relations matrix into user-item matrix factorization techniques. Specifically, a user’s ratings and the ratings of this user’s social neighbors are employed to construct a virtual user. The first social fused method exploit user latent preferences and latent perferences of this user’s social neighbors to construct social influence, the prediction error can be significantly reduced when personal factors and social influence are both intergrated into rating prediction function. However, the social influence is unique for each user: not only the number of social neighbors, but also the degree of social influence. Hence, the second social fusion model is builded by the representation of asymmetric social influence. Experimental results demonstrate the effectiveness and efficiency of both algorithms.