Deep multi-granularity graph embedding for user identity linkage across social networks

Shun Fu, Guoyin Wang, Shuyin Xia, Li Liu

Abstract: There have been increasing interests in user identity linkage (UIL) across social networks since it supports many applications such as cross-net recommendation, link prediction, and network fusion. Existing graph embedding based techniques cannot sufficiently model the higher-order structural properties in UIL. Moreover, the very limited supervisory anchor pairs (SAP), which are crucial for the task of UIL across social networks, are not utilized effectively. In this paper, a novel framework named multi-granularity graph embedding (MGGE) is proposed. And as an extension, a deep multi-granularity graph embedding model (DeepMGGE) is further developed. DeepMGGE uses the random walk (RW) to capture the higher-order structural proximities which is ignored by IONE Liu et al. (2016). Besides, DeepMGGE employs a heuristic edge-weighting mechanism given by deep learning to better capture the non-linear SAP-oriented structural properties. Experiments on real social networks demonstrate that the DeepMGGE outperforms state-of-the-art methods.