Research On User Preference-Based Recommendation Methods On Social Networks

With the popularization of computers and mobile devices,the Internet technologies have been well developed and the lifestyle of individuals has been totally changed by these technologies.The development of the Internet explodes the amount of information and causes the problem of information overload.Recommender systems can be regarded as an effective tool to solve the information overload problem,which are capable of mining useful information from the massive behavioral data of the Internet users.Recently,recommender systems not only achieve lots of attention from many academic communities,but also have been widely used in industry.Considering about recommender systems have important value in the aspects of profit growth and maintaining user viscosity,many platforms of e-commerce and social networks have developed their own recommender systems.The integration of social networks and e-commerce is getting closer.There exists abundant user behavior data on social networks and the data is an indispensable part when evaluating user preference.This thesis uses the analysis methods of complex networks and machine learning technologies to process the features of user preference on social networks in order to solve the problems of personalized recommendation and group recommendation.Many technologies,such as bipartite graphs,tripartite graphs,matrix factorization and Bayesian personalized ranking,are used to improve the performance of recommendation algorithms.The main research contents of this thesis are as follows:1.Research on mixed similarity diffusion recommendation algorithms based on bipartite graphs.Users and items in recommender systems can be abstracted as nodes on a bipartite graph and connected edges on a bipartite graph represent the users' behaviors of purchasing or collecting items.This thesis mainly researches on resource diffusion processes on bipartite graphs and proposes a recommendation algorithm based on mixed similarity diffusion.This algorithm simultaneously considers both explicit and implicit feedbacks from users and uses these two kinds of feedback data to evaluate the similarity between users.In addition,this thesis analyzes the influence of node degree on resource diffusion processes and then proposes the concept of node degree balance that uses parameters to regulate the role of various nodes in the processes of resource diffusion.2.Research on recommendation algorithms based on trust diffusion on triparitite graphs.Due to the structure limitation,it is difficult to introduce trust relationships into bipartite graphs as an additional feature for recommendation.This thesis designs triparitite graphs as an extension of biparitite graphs and propose a trust diffusion process to introduce trust relationships into recommendation algorithms.Additionally,this thesis proposes a resource diffusion process on tripartite graphs,and integrates cosine similarity between users and items into the resource diffusion process to make recommendation results more personalized.3.Research on trust-aware group recommendation with virtual coordinators.In this thesis,the concept of virtual coordinators is proposed.In the process of evaluating the personal preference of users in the group,the virtual coordinator generates global preference by observing the preference of all users in the group and interacts with users in each group to alleviate preference conflict among users by using global preference.This thesis also utilizes vital node analysis methods in complex networks to evaluate the personal influence of users using social relations.The users with high social influence can have greater impact on the virtual coordinator in interaction processes and make the virtual coordinator consider their personal preference more.Through experimental analysis,under the guidance of the virtual coordinator,the preference conflict of users in the group is significantly reduced,and the aggregated group recommendation results can better meet the needs of group members and improve the accuracy of group recommendation.4.Research on Bayesian personalized ranking for event-based social networks.Event-based social networks are an emerging form of social networks,which have more abundant social features.Aiming at the ranking problem in the recommendation domain,this thesis classifies negative samples and combines group features and label characteristics to proposes a multi-level pair-wise learning model to refines users' preference on event-based social networks.In addition,this thesis introduces the average preference of similar users into the pair-wise learning model by means of regularization term constraint.The experimental results show that the multi-level pair-wise learning model combined with social features has obvious advantages in the accuracy of recommendation results.