Research On Social Recommendation Algorithm Based On Domain Sensitivity

With the popularity of online social networks,many personalized recommendation algorithms use open behavioral data and social network buddy data to help users filter information,and find interested item more quickly.Therefore,the research of the recommended algorithms comes into being.This paper conducts a lot of research and analysis on various traditional recommendation algorithms,especially social recommendation,and finds the entry points of merges different social relationship into probabilistic models and matrix decomposition models.At the same time,we also takes into account the variation of user's interest and user's collaborative influence across different domains,and user's different willingness to be influenced.Finally,this paper proposes two novel domain-sensitive social recommendation algorithms.The main work and innovation of this paper is reflected as the following:1.Most of the trust-aware recommendation methods think that all friends with trust relationship have an impact on user ratings,which ignores the diversity of trust and the variation of user's interests across different domains.This paper proposes a social recommendation algorithm based on domain-sensitive interest circle(DSC-PMF).The DSC-PMF algorithm uses the DSC model and a probabilistic matrix factorization strategy to recommend items to users in a specific domain.The DSC model only considers the influence of friends with similar interests,and measures the influence of different friends by decomposing trust value.It also defines "the user's domain sensitivity" to assess different willingness to be influenced.2.Most of the existing domain detection algorithms does not consider the variation of the user's collaborative influence across different domain.Motivated by the observation,this paper propose a novel sensitivity domain detection social recommendation(SDD-RPM)algorithm,to make rating prediction by exploring sensitivity domain simultaneously.The SDD-RPM algorithm constructs regularization items by exploring user-item interaction information and society networks information,to build sensitivity domain detection(SDD)models,in which a domain is deemed as a circle consisting of a subset of items with similar attributes,a subset of users who have interests in these item and a subset of users with social relationships.Finally,the proposed unified framework of SDD-RPM exchanges the sensitivity domain information into predictive score by utilizing user's and item's latent feature vector,which guides the exploration of the potential feature space of the user and item by sensitivity domain information.3.The proposed algorithm is valid by designing experiment on the real datasets Yelp and Epinion for recommendation accuracy.The experiment results show that our proposed algorithms outperforms the current popular recommendation algorithms in improving the prediction accuracy,alleviates the sparsity and cold start problems at some extent.