Meta-Path Based Anchor Link Inference in Multiple Partially Aligned Social Networks

To enjoy fantastic services or fulfill their multifarious needs, people usually get involved in multiple online social networks at the same time. Due to the protection of privacy and the separation of different networks, the multiple accounts of the same user in different social networks are generally isolated. Finding the correspondence between users can benefit a variety of applications, such as cross-network recommendation and cold start. In this paper, we focus on inferring the correspondence between user pairs across partially aligned networks where some common users are shared, which is formally defined as multi-network anchor link inference problem. To solve the problem, an integrated anchor link inference framework, MALI (multi-network anchor link inference) is proposed, which consists of two parts: firstly, building structure-rich probabilistic networks with PU link prediction model to against noises. Secondly, taking advantage of cross-network social meta path as a powerful tool to extract useful features to train a classifier. Then by incorporating username with the ranking list which is based on the output probability of the classifier, an improved stable match model is designed to infer anchor links which are restricted to one-to-one relationships. Extensive experiments on two real-world partially aligned heterogeneous networks show that our proposed MALI can solve the multi-network anchor link inference problem very well.