Research On User Interest Modeling And Recommendation Technology Based On Network Community

With the rapid development of information technology, burgeoning social networks providefull equipped communication platforms for human interaction, information transmission, andknowledge sharing, which enforce a significant influence on people’s life and behavior. User onsocial networks can participate in the creation and dissemination process of information anytimeand anywhere, resulting in rapid growth of information resources. How to effectively organizeand analyze the grand large social networks, accurately obtain users’ interests preference, andrecommend high-quality personalized information to user according to his or her individualinformation requirements have become hot research fields of data mining. This dissertationintroduces network community detection method into user interests modeling process, researcheshow to find community unit on social networks effectively, model group interests based on socialtagging data in the context of community, and recommend high-quality information forindividuals based on user relationship mining and analysis within community. Researchcontributions of this thesis are listed as follows:(1) Traditional community detection algorithms based on non-negative matrix factorizationdetermine the number of communities by modularity optimization, which result in highcomputational costs and suffer resolution limit. On account of this, an overlapping communitydetection algorithm based on Bayesian non-negative matrix factorization is presented. Firstly,performances of network feature matrices in community detection are analyzed, and theexpansion of adjacency matrix is selected as the input of the proposed algorithm. Then, Bayesiannon-negative matrix factorization model is introduced to calculate the number of communitiesiteratively with high efficiency. Finally, membership index between nodes and communities aredefined, and overlapping community structures are obtained by setting a reasonable partitioningthreshold. Experiments on computer-generated networks and real social networks show that thenew algorithm can effectively reveal overlapping community structures in social networks.(2) To avoid data sparsity and limitations of traditional interest modeling method, a newuser group interest modeling algorithm based on social annotation techniques is proposed bytaking the unique advantages of social tags. Firstly, a method combining tag recommendationand new tag extraction is employed to add tags to unlabeled documents automatically. Then,semantic properties of tags are exploited to model the “point of interest”(POI) of user group, aWordNet-based semantic similarity measure of social tags is defined, and spectral clusteringalgorithm is utilized to merge different classes of users’“point of interest”. Finally, tag sets withcategory structures are represented as POIVector, filling the role to describe interests of the whole community group. Experimental results confirm that the proposed algorithm can clustertags semantically in a multi-user existing data circumstance, and the adoption of POIVector todescribe community group interests presents sound effects.(3) Traditional personalized recommendation methods usually only pay attention toimproving the accuracy of recommendation, and neglect the influence of recommendationdiversity on user’ experience. On account of this, user relationship mining techniques withincommunity are applied to personalized recommendation, and a multi-strategy recommendationalgorithm based on user relation mining is proposed. Firstly, trust propagation model is used tomine trust relations between users, similar relations between users are calculated based on cosinesimilarity of user profile, then, favoring items of target user’s “trusted friends” and “similarfriends” are taken as recommendation data candidates. Secondly, four combination strategies ofusers’ trust and similar relationship are discussed, deriving a comprehensive correlation measurebetween users, and a collaborative preference prediction mechanism is proposed to calculatepreference scores of item candidates. Finally, recommendation list is generated based onpreference predictive value of items, and Top N principle is exploited to selectrecommendation items for target user. Experimental results confirm that the proposed algorithmnot only reduces the influence of data sparsity, but also balances recommendation accuracy withdiversity indicators tactfully, which improves the overall performance of recommendationalgorithm.