Research On Recommendation Algorithm Using Context And Trust Relationship

With the development of Internet, network information growing rapidly, which brings the problem of information overload, it becomes harder for people to obtain valuable information. Personalized recommendation system can effectively solve the problem, so it has received more and more attention. With the rise and popularization of network community and personal intelligent mobile terminal, more valuable information, like the context and trust relationship, can be used in recommendation system. Trust relationship can enhance the accuracy of recommendation algorithm, and the context information can make recommendation system more personalized, if we try to use both of them in a better way, we can provide users with higher quality recommendations. The researchers have done a lot of research on trust-based recommendation system and context-based recommendation system, but the research on how to use context and trust relationship simultaneously is relatively less. The data sets of recommendation system usually contain tens of millions of information from both users and items, how to improve efficiency of recommendation algorithm for huge amounts of data is also another important part of this thesis.Therefore, in this thesis, we do research about the recommendation algorithms using context and trust relationship simultaneously, and parallelize the newly developed algorithms in MapReduce framework. The main work is as follows:1. Propose two recommendation algorithms, named SocialMF and Trust-FMs, which use context and trust relationship simultaneously, and learn the Trust value matrix dynamically to further improve the quality of recommendation.(1) Context-SocialMF is based on contextual post-filtering, the main idea is:1) Use SocialMF to get the preliminary recommended result sets; 2) Use context preference extraction algorithm to extract the user’s contextual preference; 3) Use user’s contextual preference to filter the preliminary recommended results by nearly location priority strategy. Trust-FMs restructures each user’s feature by their friends’, regards the trust value between users and close friends as the weight, thus the trust relationship is joined in the contextual recommendation algorithm FMs.(2) The above two algorithms use the given trust values in advance, which are difficult to describe the real trust relationship between users accurately. To this end, this thesis builds optimization function of the trust value matrix, and solves optimization function by using gradient descent method to learn trust relationship dynamically. So that trust value matrix of users’ relationship can reflect the strength of trust between users, hence recommendation quality can be effectively improved.2. Research on how to parallelize the Context-SocialMF and Trust-FMs in MapReduce framework.The Context-SocialMF and Trust-FMs belong to matrix factorization recommendation algorithm, reducing the size of the matrix is an important strategy to improve the efficiency of the algorithms. Therefore, in this thesis, we devide the matrix to vectors, all vectors are evenly distributed to each node, so that each node can use different vectors into do computing at the same time. In this way we achieve the goal of parallelizing the algorithms in MapReduce framework.