A Study Of Collaborative Filtering Based On Sequential Behaviors Mining And Privacy Protection

With the increasing prosperity of Web2.0, more and more users and applications are emerging on the Internet, and their interactions generate a large number of data. The rapid expansion of data leads to the serious problem of information overload. As a result, people cannot obtain information to satisfy their personal needs in a short time. Under this background, various recommendation systems are developed to capture the personalized information needs of each single user through learning his/her preferences. In this way, recommendation systems can effectively solve the problem of information overload, and thus have been widely studied and used. However, in recent years, the rapid development of the Internet brings in many new research challenges to the development of recommendation systems.To that end, this thesis first describes the information overload problem, and gives some preliminaries of recommendation systems as well as some state-of-the-art recommendation algorithms. Then, this thesis provides a focused study on solving the problems of sequential behaviors mining and privacy protection in collaborative filtering based recommendation systems. Specifically, the major contributions can be summarized as follows:(1) This thesis proposes a novel method to capture the sequential behaviors of users. Specially, this method can mine the set of neighbors that are most influential to the given users and items. Through building the consumer network, it calculates the users’nearest neighbor set effectively. Compared with social relationships and tag information, the sequential behaviors are easier to be obtained, which makes this method more applicable.(2) This thesis applies the mined sequential behaviors into the matrix factorization model and develops the SequentialMF model. This model integrates the mined nearest neighbors set into probabilistic matrix factorization algorithm to improve the recommendation accuracy. Moreover, based on this model, this thesis further proposes a recommendation framework which can improve real-time performance of recommendation systems.(3) This thesis proposes a recommendation mechanism named GroupMF which can protect the privacy of online users in the recommendation system. Specifically, it uses the group-level information of users to achieve personalized recommendation. As this mechanism does not require the users’personal preferences on a specific item, it can achieve the goal of privacy protection.(4) This thesis validates the effectiveness of the proposed models by extensive experiments. Experimental results show that1) the recommendation accuracy of the SequentialMF model outperforms conventional algorithms.2) GrouMF has the good ability to protect the privacy of the users with little loss of recommendation performance. In addition, the input data for GroupMF can be easily collected with a high quality.