Research On Personalized Recommendation Methods For Users In Online Social Network

As an important technique to deal with information overload,personalized recommender systems have been widely applied into various electronic commerce platforms such as Amazon and JD.com to improve users' browsing experience.The proliferation of online social network brings rich data source such as various types of user connections in online social network like trust or membership,providing sufficient fuel for the recommender engine,along with more recommendation scenarios.It has been the research focus to incorporate user connections into recommender systems for both industry and academia.The sparsity of user connections,the different role among various user connections and the complexity of interaction between users and communities in group activities bring new challenge for user profile modelling in online social network.It is significant to propose appropriate recommendation methods for online social network users because this not only enriches the personalized marketing theory and improves the marketing efficiency for companies.This thesis focuses on the personalized recommendation methods for online social network users and employs the machine learning algorithms to investigate the recommendation techniques based on explicit or implicit user relationships and group recommendation approaches in online virtual communities.The detailed research content are summarized as follows:(1)Personalized recommendation method based on explicit user relationships.To address the sparsity of user connections and trust or distrust between users,in Chapter 2 we propose a trust prediction model based on weighted social trust.This trust prediction model transform the unweight trust or distrust relationships into weighted trust relationships,which can capture fine-grained trust or distrust degree between users.Then we develop a collaborative matrix factorization model for personalized recommendation which combine user trust relationship and user item interaction.Empirical studies on Epinions dataset show that the proposed recommender that considers user trust relationships can improve the recommendation accuracy,especially under the extreme sparse data or cold start user scenarios.(2)Personalized recommendation method based on implicit user relationships.To differentiate the role of implicit connection between users on user decision making,in Chapter 3 we propose a Social Impact based Latent Dirichlet Allocation model(SILDA-1)for personalized recommendation.Specifically,SILDA-1 model employs a switch variable to differentiate the role of social impact and personal interest during the decision making process.In addition,considering the fact that different users has different social impact in online social network,an extended model based on SILDA-1,SILDA-2 model is proposed.SILDA-2 model can automatically learn the social impact weight for each individual.Experimental results on CiteULike dataset demonstrates the effectiveness of the proposed SILDA-1 and SILDA-2 model,especially under the context of extreme sparse data and cold start user scenarios.(3)Personalized recommendation methods for group of users.To investigate the influence of group on user interest,in Chapter 4 we propose a Bidirectional Tensor Factorization model for Group Recommendation(BTF-GR),which can capture the interaction between individual's intrinsic interest and group impact.Furthermore,we utilize personalized weight to represent the group impact difference among different users.Empirical studies on two real-world data sets,CiteULike and Last.fm,demonstrate that the proposed model outperforms the baseline algorithms,especially for homogeneous groups.The innovation points of this research are listed as follows:(1)Category relationships between users in social network into explicit connections and implicit connections.(2)Propose a personalized recommendation method by utilizing explicit connections between users where the connections are extremely sparse.(3)Propose a personalized recommendation method by utilizing implicit connections between users,which can differentiate the user personal interest and the social impact.(4)Propose to model the group recommendation as a bidirectional procedure and firstly apply tensor factorization model into group recommendation problem.This paper expands the research area of the personalized recommendation techniques,enriching the research content for the data mining and social network analysis,and provides several recommendation algorithms under various recommendation scenarios.