Learning user distance from multiple social networks

In this paper, we propose an adaptive user distance measurement model to address the challenging problem of modeling user distance from multiple social networks. Previous works construct user distance model in a single social network, and dataset easily leads to over-fitting of the models due to the data sparseness of a single sparse network. We observe that people often simultaneously appear in multiple social networks, because different social networks (e.g., Facebook, LinkedIn, QQ, Douban, etc.) can provide complementary services. Thus, the knowledge from different social networks can help overcome the problem of data sparseness. However, knowledge cannot be directly obtained due to that it is from different social networks. Aiming to solve this problem, we construct an adaptive model to measure user distance from multiple social networks by employing the metric learning and the boosting technology. The basic idea of our model is to embed multiple networks into a potential feature space while retaining the topology of social networks. In the procedure of boosting, the negative effects caused by network differences and useless information can be avoided. To get the solution of our model, we formulate it as a convex optimization problem. Besides, we propose an Adaptive User Distance Measurement (AUDM) algorithm whose time complexity is linear to the number of the links. Finally, we verify the feasibility and effectiveness of AUDM on the problem of link prediction. Experiments on a real large-scale dataset show that AUDM outperforms the state-of-the-art algorithm.