An Approach for Building Efficient and Accurate Social Recommender Systems using Individual Relationship Networks (Oct-2017)
The main aim of this project is to enhance the social recommender system using social network and improve the accuracy of traditional recommender systems.
We propose the best approach to address the infused complexity of social relationships in social recommender systems. The approach can be applied to both user-user and item-item relationships. We define the individual relationship network (IRN) for each user or item, and present an algorithm based on the similarity, density and confidence measures to make a balance between its accuracy and efficiency. The recommendation method addresses the cold start problem, captures the diversity of tastes between connected users, and enables scalability by fusing MF and neighborhood models via IRN’s. Our method generates an individual relationship network for each user/item to control the complexity of social relations. Moreover, existing social recommendation methods try to reduce data sparsity and cold-start users from the user perspective, but the cold-start problem for items still remains. Our approach incorporates item relationship network by using user-oriented and item-oriented perspective and can address the cold-start problem for items.
We fuse the neighborhood model and MF via IRN’s to maximize the potential of relationship network. Unlike the regularization method with predefined similarities in relationship network, our method models the taste diversity between relationship members as dynamic similarity constraint to capture the time-evolving nature of tastes in model learning. Unlike earlier works with a manual control of balance coefficient, our method targets the social influence as an extra user-item specific bias and absorbs the balance coefficient into an interpolation weight matrix which represents the influence a user exert on another user, since the influence is learned from the data automatically.