REBORN: Transfer learning based social network alignment

Abstract

Social network alignment (SNA) aims to identify the same person across different social networks. Existing SNA methods can be categorized as supervised, semi-supervised and unsupervised. However, for supervised and semi-supervised methods, obtaining supervisions is expensive and time-consuming; for unsupervised methods, their performances will reach to a bottleneck due to the lack of supervisions. In the real world, there exist many social networks which have already been aligned spending money and time, and many social networks to be aligned. To bridge this gap, we study the SNA problem from the transfer learning perspective, and propose to leverage the knowledge of two aligned social networks (source domain) to facilitate the alignment of two unaligned social networks (target domain). Specifically, we first propose Ego-Transformer to align two social networks in the same domain. Then, we propose WWGAN to eliminate the shift between different domains. Finally, we propose REBORN which tactfully incorporates Ego-Transformer and WWGAN together. We conduct extensive experiments on the datasets of Facebook-Twitter and Weibo-Douban, and compare REBORN with 8 state-of-the-art SNA methods including SNNA, MSUIL, BRIGHT, DeepLink, IONE, PALE, WAlign, and UUILgan. The results demonstrate that REBORN can averagely achieve 20.02% higher Precision@k and 23.23% higher MAP@k compared with these methods.