GRANA: Graph convolutional network based network representation learning method for attributed network alignment

Abstract

Social network alignment, which aims at identifying the correspondences of the same users across networks, is the very first step of information process from multiple social networks. Previous efforts on this task are either more inclined to preserve structural consistency or attribute consistency. Therefore, they only achieve good performance on specific alignment tasks or obtain compromised results on all kinds of alignment tasks. To achieve good generalization, in this paper, we propose a novel multi-task learning method to solve different social network alignment tasks, which is named GRANA (Graph convolutional network-based network Representation learning framework for Attributed Network Alignment). Specifically, a new two-layer cross-network convolutional neural network dubbed Cross-GCN is proposed as shared layers of GRANA. And the intra-network and inter-network attribute and structural information are learned respectively with diverse objective functions in the task specific layer of GRANA. To enhance the alignment performance and accelerate the learning process, a weight learning method with a novel weight initialization process is applied. Experimental results on six kinds of datasets show that GRANA outperforms seven state-of-the-art methods by at least 0.002-0.697 in terms of precision@15 value. The ablation studies further support the effectiveness of proposed Cross-GCN and weight initialization process.