Research On Personalized Recommendation Methods By Fusing Multiple Influence Factors In Social Networks

With the continous development of computing socialization,recommender systems(RSs)have been widely used in e-ecommerce,mobile news,social networks,and etc.As an important kind of tools for filtering information,RSs aim at recommending information actively for users to meet their personalized needs from massive data.However,traditional RSs only use user's historical behavior data,namely the binary relationship between users and items to model user's preferences and interests.In the process of recommendation,it is difficult for recommendation quality to meet user's needs due to such problems as data sparsity,cold start,scalability,and so on.In recent years,with the rapid development of online social networks,kinds of social relationship data provides a rich data source for personalized RSs,which brings an opportunity to solve the problems of data sparity and cold start.At the same time,the sparsity and complexity of the connections of social relationships between users in social networks and the difference of the influence of social relationships on users' decisions have become new challenges to model users' interests in social RSs.How to integrate the influences of social relationships in the recommendation algorithm and improve the performance of the recommendation has become one of research hotspots in the field of RSs.The existing recommendation methods mainly use the direct trust relationship to model user's interests and preferences from the technical point view,while ignoring the influences of indirect social relationships on recommendation performance.In this dissertation,we focus on the data sparsity and cold start in RSs,analyze the impacts of kinds of interactive information on recommendation performance from the perspective of sociology,use explicit and implicit social relationships to model user's interests,and study the social recommendation method integrating various influence factors.The main work and contributions of this dissertation are described as follows:1.A hybrid recommendation approach based on user's multi-interest mining is proposed.Data sparsity is an important factor affecting the clustering effect and recommendation quality.For the extremely sparse problem of ratings on items from users,the hybrid recommendation approach implements the prediction of the missing ratings and make a recommendation in the following two steps: Firstly,from the perspective of improving the quality of clustering,the user-item rating information is converted into the user's preference for item's attributes by combining the item's attribute relationship to alleviate the data sparsity problem,the users are clustered according to the user's interests based on the probabilistic latent semantic analysis model,and the user's multi-interest preferences are mined to pre-fill the partial missing ratings.Secondly,an enhanced social matrix factorization model is proposed by combining user's trust relationship,item's similarity and matrix factorization technology to predict all missing ratings.The experimental results demonstrate that the proposed hybrid recommendation approach outperforms the existing social recommendation methods in the accuracy and diversity of the recommendation.2.A novel recommendation approach by fusing user's social status and homophily is proposed.The trust relationship between users reflects the degree of interdependence between users in social networks.The strength of trust between users depends on their degree of expertise in a certain field.The same user have different degrees of expertise in different fields,that is,different authoritative.Therefore,the contribution of each user to trusted users in different fields should be different.At the same time,users with the same social background are likely to have similar interests and preferences.However,the current recommendation algorithms treat trusted users equally in social networks,ignoring the impacts of the difference in the degree of trust of users in different fields and the interest consistency of users with the same social background.To solve the above problems,considering the difference of influence on their trusted users from users with different degrees of authority in their respective fields and similarity of preferences of users in similar social backgrounds,this dissertation analyzes the influences of user's social status and homophily on recommendation quality from the perspective of sociology,establishes corresponding weighted relationship models,and integrates them into the matrix factorization model to improve the quality of the recommendation.3.A recommendation approach based on social relationship context using social interaction factors is proposed.The context of social relationships,that is,the connections between users in social networks,plays an important role in improving the quality of the recommendation.The interaction information between users and items in social networks reflects indirectly the social relationships between users.In view of the sparseness of social relationships,this dissertation deeply analyzes the social relationships implied in the interactions between users,between users and items,and between item attributes,enhances the social relationships between users and between items,and integrates the improved social relationship influence factors into the matrix factorization model.At the same time,the user's social relationship and the user's interest similarity are used as regularization terms to constrain the user's latent feature space to approach the real value.The experimental results demonstrate that the accuracy of the proposed approach is improved compared with the current social network recommendation algorithm.4.A social recommendation approach based on implicit similarity by fusing trust relationships between users and social tags is proposed.The social relationships between users in social networks are complex and diverse,so direct modeling of social relationships may easily lead to the distortion of user's preference model,which will affect recommendation performances of RSs.To solve the problem of low recommendation accuracy caused by the directly measuring the social relationship between users and between items,an improved social recommendation approach that integrates trust relationships and social tags is proposed.By analyzing the impacts of explicit and implicit social relationships on the recommendation quality,user latent feature and item latent feature are mapped to a shared space,respectively,and they are trained continuously to obtain more accurate similarity relationships between users and between items using multiple influence factors,thereby the performance of the recommendation is improved.The experimental results demonstrate that the proposed approach outperforms other social recommendation approach in the recommendation accuracy and diversity,and verifies the rationality of the idea objectively.