Font Size: a A A

Research On Social Relation Based Recommendation Algorithms In Social Networks

Posted on:2016-03-12Degree:DoctorType:Dissertation
Country:ChinaCandidate:L GuoFull Text:PDF
GTID:1108330461985404Subject:Computer system architecture
Abstract/Summary:Request the full-text of this thesis
Recommendation methods as one of the important information filtering techniques have been well used for solving the information overload problem. The main task of these methods is to suggest users the potential items of their interests inferred from their history ratings and behaviors. Successful applications of recommendation methods can be found in many industry areas, such as e-commerce, movie & video sites and Computational Advertising. However, traditional recommendation methods ignore the social relations among users, which are important for us to make the right decisions. For example, in our real life, we always turn to our friends for recommendations. The friend relationship guarantees the confidence level of their recommendations. Hence, with the advent of online social networks (e.g., Twitter, Facebook), the social relation based recommendation methods have attracted a lot of attentions.Based on the existing problems and challenges, this thesis investigates the social relation based recommendation methods in social networks under the support of NSF. The main research contents and innovations of this thesis, includes a recommendation method incorporating item relations, a trust strength aware social recommendation method, a recommendation method with social contextual information and a trust aware social item ranking method:1. We propose a recommendation method incorporating item relations.In social networks, item relations are important factors that can help us understand users and model their interests. In this work, we introduce the item relations to the recommendation problem and utilize them to constrain the objective function in a shared latent feature space. In this way, in our method can consider the influence of user connections, user interests and item relations simultaneously. Experimental results in real social network show that the proposed approach outperforms the recommendation methods without item relations in terms of precision and rating error.2. We propose a trust strength aware social recommendation method.Although trust relations are important to model user interests, most existing trust aware recommendation methods treat trust relations equally and assume trusted users must have similar tastes. But in fact, two users may establish a trust connection for the social purpose or simply for etiquette, which may not result in similar opinions on the same item. Motivated by this observation, a trust strength aware social recommendation method is proposed. This method learns the trust strength and distinguishes users with more similar interests through the shared user latent feature space. To validate the learned trust strength, we take SocialMF as an example, and retrain it with estimated trust relations. Experimental results show that the trust strength is one of the important factors to reflect user interests, and our proposed approach can achieve better performance than baseline methods.3. We propose a recommendation method with social contextual information.Contextual information is important for us to understand user behaviors and help them make the right choices. Contrary to traditional context, in social recommendation scenario, social context is the user link status of her social network. Based on the above observation, we incorporate social contextual information into the recommendation problem, and use this information to extend the user latent feature vector. To further improve recommendation quality, we utilize the common social relations as factorization terms to regularize the ranking objective function. Experimental results demonstrate that the social context is the important information to model user interest, and our approach can outperform the other state-of-the-art algorithms.4. We propose a trust aware social item ranking method.Although trust relations have been well used in rating based recommendation methods, few methods have studied the function of these information in ranking based recommendation tasks. In this thesis, we take the recommendation problem with only implicit feedback as an example, and propose a trust aware social item ranking method to investigate the influence of trust relations. Specifically, we first derive a social based personalized item ranking criterion for the implicit feedback from a Bayesian analysis of the problem, where we introduce the social trust assumption from the view of social influence theory to improve the item recommendation performance. To explore the impact of user multi-faceted trust relations, we further propose a category-sensitive random walk method to infer the true trust value on each trust link. Data analysis and experimental results on two real-world datasets demonstrate the existence of social trust influence and the effectiveness of our social based ranking methods.
Keywords/Search Tags:Recommender System, Social Network, Matrix Factorization, Social Influence, Learning to Rank, Collaborative Filtering
Request the full-text of this thesis
Related items