Research On Collaborative Filtering Recommender Method Combining Social Relationships Of Mobile Users

With the development of mobile internet brings information overload. Recommender system has attracted focus in mobile domain as a method to effectively alleviate information overload. It is indispensible for user-centered services to provide personalized services for users with recommender methods. Collaborative filtering techniques are widely implemented in recommender systems. The basic idea of collaborative filtering is to make recommendations for the active user with other similar users’ information. Due to social influence in social network, users with relationship tend to have similar preferences. Social relationships of users are incorporated in recommender method to improve recommendation performance. In this thesis, some key problems of collaborative filtering recommendation combining with social relationships of mobile users are studied. We summarize our main contributions as follows:(1) A collaborative filtering recommender method combining item feature and trust relationship of mobile users is proposed. User-based collaborative filtering recommender method exploits preferences of similar users to predict the active user’s preferences for unrated items. Usually similarity between two users is calculated based on the users’ preferences of common rated items. Items rated only by one user would not be exploited to calculate the similarity between two users. Especially when users’ ratings are sparse, the number of common rated items between users becomes small. Similarity between users can’t be precisely measured on small number of items. Based on this, this paper exploits EMD (earth mover’s distance) method to calculate similarity between mobile users. The method exploits mobile users’preferences on common rated items and also incorporates mobile users’preferences on similar items. And in preference prediction of collaborative fitering recommendation, we propose to predict preferences with similar users’ preferences of similar items. To alleviate the influence of data sparsity, preferences of trust users and similar users are combined to predict user’s preferences for unrated items. The experimental results show that similarity between users calculated with users’preferences on similar items, preference prediction with users’ preferences on similar items and with preferences of trust users can improve prediction accuracy and recommendation accuracy of collaborative filtering recommender method.(2) A collaborative filtering recommender method combining with trust relationship of mobile users is proposed. Due to social influence between trust users, performances of collaborative filtering recommender system can be improved by incorporating trust relationship of users. The accuracy of social influence between users would determine performances of recommender systems. Methods based on preference similarity of users can’t precisely measure social influence between trust users. Because the preference similarity is calculated based on common rated items between users and the similarity between users is symmetric. But social influence between users is not symmetric. Based on this, we exploit the mechanism of information diffusion and social influence between trust users to select influence item set between two users. Meantime, to alleviate influence of data sparsity, implicit preferences of mobile users to unrated items are predicted with their trust relationship. And implicit preferences are used to select influence items. Then the social influence between trust users will be calculated in influence items. To evaluate the proposed method, experiments are conducted on public datasets. The experimental results show that the proposed method can get better accuracy of prediction.(3) Hybrid methods combining similar relationship and trust relationship of mobile users are proposed. Traditional collaborative filtering method exploits preferences of similar users to make prediction. And collaborative filtering method with trust relationship of mobile users exploits influence of trust users to make prediction. Similar relationship and trust relationship of mobile users are favorable to improve performance of collaborative filtering method. But in recommendation, similar relationship or trust relationship is neglected. To improve performances of recommender system, feature-based hybrid method, cascade-based hybrid method, meta-based hybrid method and meta-based hybrid with cascade-based hybrid method are proposed, which are based on similar relationship of mobile users and trust relationship of mobile users. Performances of hybrid methods are evaluated on public datasets. The experimental results show that combining similar relationship of mobile users and trust relationship of mobile users can improve performances of recommender system.(4) A group recommender method combining with group relationship of mobile users is proposed. Currently group recommendations are divived into aggregated model method and aggregated prediction method, which include predicting step and aggregating step. Group recommender method predicts group members’ preferences with recommender technique and then aggregates group members’preferences into the group preferences with aggregation strategy, or aggregates group members’ preferences into preferences of groups with aggregation strategy and then makes prediction for groups based on groups’preferences. Aggregated model method exploits preferences of group to predict preferences of groups and make recommendations, which can easily cause loss of members’preferences. And aggregated prediction method neglects user’s group relationship when predicting members’preferences. To solve the problem in group recommender method, we propose a group recommender method based on matrix factorization. The method combines users’group relationship with matrix factorization and exploits users’group relation and members’ preferences to learn groups’preferences. And the influence of group members on group preferences is different. Hence, the weight of influence of every group member is combined with the matrix factorization method. The proposed method is compared with aggregated model and aggregated prediction group recommender methods. And the experimental results show that the proposed method can improve performances of group recommendation.