Research On Distributed Group Recommendation Algorithm

With the development of information technology,the data on the internet is on exponential growth.In the era of big data,the traditional stand-alone recommendation algorithm is hard to recommend users the commodities which they are interested or needed in a short time.Therefor,the distributed recommendation algorithm has been studied.Meanwhile,in recent years with the rapid development of social network,such as Weibo in China,it has become an indispensable part of people's daily life.Due to people's frequent social activities,Weibo users can often be divided into several groups.Recommending a friend to a group,so called group recommendation,is appeared.However,facing with such a huge amount of Weibo data,how to analyze it with data mining,machine learning and other methods,especially how to recommend friends to users has become a hot research direction in the field of social recommendation.Through studying a parallel matrix decomposition on a distributed programming platform,this thesis proposes a recommendation algorithm based on matrix LU decomposition,called LUALS-WR that can more efficiently complete the recommendation for a large number of data while there is no significant reduction in recommended accuracy.Moreover,after applying the proposed LUALS-WR algorithm for recommending friends to individual users,this thesis proposes a grouping strategy based on social choice for Weibo user group recommendation.The main work is as follows:(1)This thesis analyze and study the performance of traditional collaborative filtering recommendation methods based on matrix decomposition,such as SVD and SVD++.Experiments are conducted on a Tencent Weibo user-to-user relationship data.(2)This thesis proposes a LUALS-WR algorithm to accelerate matrix updating by using LU Decomposition.The original parallelization matrix decomposition algorithm(WR-ALS)in the MapReduce framework has been improved by the proposed algorithm.The key point is that LU decomposition is used to complete the matrix inversion instead of QR decomposition,thus making the fast update rate of the feature matrix.(3)This thesis analyzes the traditional fusion strategies,and proposes a group preference fusion strategy based on social choice for group recommendation.Compared with traditional fusion strategies,the proposed method can obtain higher group recommendation precision in terms of RMSE.(4)This thesis conducts a comprehensive experiments to verify the proposed methods with the 3.8G Tencent microblogging data.Experimental results show that the LUALS-WR algorithm is 1.5 times faster than WR-ALS algorithm while their recommendation precision is almost the same.Morevoer,the recommendation precision of the proposed social preference-based fusion method is about 5% higher than the traditional preference-based fusion method in group recommendation.