A Multi-Granularity Representation Learning Framework for User Identification Across Social Networks

Abstract

Predicting latent user identity across social networks has many application scenarios and it can be demonstrated by using the similarity of network structure caused by the similarity of friend relationships. Former related research works consider the structural similarities only while not sufficiently modeling the higher order structural properties. Moreover, the very limited supervisory anchor pairs, which are crucial for the task of user identification across social networks, are not utilized effectively. Based on the idea of multi-granularity cognitive computing and for partly solving the problem of multiple granularity representation of data proposed in DGCC (Data-driven granular cognitive computing) [23], this paper proposes a high-performance framework called multi-granularity representation learning (MGRL) framework for user identification across social networks which facilitates a well-designed heuristic mechanism to weight the edges on which a guided sampling strategy is conducted for vertex sequence generation. This enhances the model’s capability of capturing the higher-order structural proximity. By integrating two aspects of structural properties, the multi-granularity structural features are preserved well. Experiments on real life social networks demonstrate that the MGRL significantly outperforms other state-of-the-art methods on the task of identifying latent corresponding users across social networks.