Heterogeneous Networks Based Random Walk Without Meta-paths For Social Recommendation

Social recommender systems can alleviate the problems of data sparsity and cold-start,which has been widely concerned by researchers and industry.However,social relations are also sparse and meanwhile are usually noisy.Therefore,a few studies identify more reliable implicit relations for each user over the user-item and user-user heterogeneous information networks.Among these research efforts,meta-paths guided search shows state-of-the-art performance.However,designing meta-paths requires prior knowledge from domain experts,which may hinder the applicability of this line of research.To solve this problem,we propose a social recommendation model with a meta-path-free strategy.Concretely,we adopt the idea of ‘Jump and Stay' in heterogeneous random walk to identify implicit friends to recommend,and form the social recommendation model with meta-path-free strategy.Then we divide the user social network into communities to form the social recommendation integrating user group information with meta-path free strategy.Experiments on two real-world datasets show the superiority and feasibility of the proposed method.The main work of this thesis is as follows:(1)The thesis introduces the development of social recommender systems.We analyze the current research status of social recommender systems,and analyze the advantages and disadvantages of current social recommender systems,according to the trending of technology development and related models,summarize meta-paths and heterogeneous information network embedding technology.(2)We build a heterogeneous information network composed of the positive and negative feedback information of users to item,and social network.Then leverage the idea of jump and stay to walk randomly in the formed heterogeneous information network to obtain a series of node sequence corpora for each node in the network structure.Then,the embedding of the corresponding nodes is learned by the heterogeneous information network representation learning method.Then we calculate the vector similarity between users to identify the implicit relationship.The implicit relationship is combined with the enhanced social Bayesian Personalized Ranking model to generate the corresponding personalized recommendation list for each user.(3)We leverage the community discovery technology to divide user social networks into groups as virtual node domains.Then we build a heterogeneous information network composed of the virtual node domains,the positive and negative feedback information of users to item,the social network information,adopting the idea of ‘Jump and Stay',to complete the social recommendation task of integrating of user group information.(4)We compare the overall performance of the proposed methods in this thesis and analyze the influence of related parameters on two real world datasets Last FM and Douban.The results prove the efficiency and feasibility of the proposed methods.Meanwhile,we compare the implicit relationship based on meta-paths and that based on jump and stay and demonstrate the reason.