Research On Recommendation Methods In Social Networks

Online social networks are becoming the most popular platforms of information exchange, where users build friendship connections and share their interest information. Some popular social network sites such as Facebook, Twitter, and Sina Weibo, are attracting thousands of new users each day. The amount of user-generated information is increasing far more quickly that users can’t handle the information overload without the support of recommendation methods.Collaborative filtering based recommendation methods are one of the most successful solutions to the information overload issue, which have been widely used in real world recommendation services and systems. Generally, collaborative filtering based recommendation methods build on the user-item similarity measures. The basic idea underlying collaborative filtering methods is that if two users have historically had similar interests on some items, they are likely to be interested in other items similarly. However, collaborative filtering based recommendation methods treat user-item information equally and ignore the context information in social networks, such as social relationship, social influence, information propagation over the connections between users and temporal factor, which affect the qulity of recommendation methods.Aiming at improving the quality of recommendation methods in social networks, this thesis addresses a number of important questions regarding the performance of recommendation, and then proposes the recommendation methods. The main contributions of this thesis are following.1. It proposes a social item recommendation (SIR) model based on neighborhood model and latent variable model. The SIR model encodes both the interest and friendship information, associates the latent variables with the interest similarities between the pair of the active user and his followers. Then extend the SIR model to SIR+for considering the social features during the inference of social item recommendation. The experimental results demonstrate that both SIR and SIR+outperform the traditional collaborative filtering methods, and SIR+achieves a better performance than SIR. 2. In order to address the information propagation impact on recommendations in social networks, it applies a user interest model (UInRec), which characterizes the user interests and interest propagation in online social networks. UInRec fuses both social network features and user-item click information for recommendations. Firstly, UInRec uses collaborative filtering techniques to model the user interest from user-item click information. Secondly, UInRec models the strength of followship by social action features. UInRec combines both the user interest model and interest propagation model for social recommendation and improves the performance of recommendation in social networks.3. Considering the trust relationship in social networks for recommendation, it presents a trust metric to quantitatively measure the recommendation trust between pairs of users by aggregating the implicit trust and trust propagation values, and then selects the neighbours based on the trust metric. After that, it proposes a trust-based latent factor model, which allows us to incorporate pairwise trust values into the latent factor model for top-k item recommendation. Finally, the experiments are conducted on Sina Weibo and the results show that the proposed method leads to a substantial increase in the performance of top-k item recommendation.4. Exploiting the context information such as time, relationship and user feedback information in social network, it presents a time-aware social recommendation method based on user feedback for top-k item recommendation. This method incorporates the temporal factors by introducing a time weight function, which models the decay of user interest. Moreover, the method considers the user positive feedback and negative feedback information, as well as the social relationship information for recommendation, Experimental results and analysis show that the proposed method outperforms the collaborative filtering method for top-k item recommendation in social networks.