Cross-Graph Representation Learning for Unsupervised Graph Alignment

As a crucial prerequisite for graph mining, graph alignment aims to find node correspondences across multiple correlated graphs. The main difficulty of graph alignment lies in how to seamlessly bridge multiple graphs with distinct topology structures and attribute distributions. A vast majority of earlier efforts tackle this problem based on alignment consistency, which directly measures the attribute and structure similarity of nodes. However, alignment consistency is prone to be violated due to the radically different patterns owned by different graphs. Another group of methods tackle the problem in a supervised manner by learning a mapping function that maps the node representations of both the source and target graphs into the same feature space. However, these methods heavily rely on observed anchor links between different graphs while these anchor links are usually limited or even absent in many real-world applications. To address these issues, we propose an unsupervised cross-graph representation learning framework to jointly learn the node representations of different graphs in a unified deep model. Specifically, we employ an auto-encoder model to learn the cross-graph node representations based on both attribute and structure reconstruction, where source and target graphs share the same encoder but are decoded by their respective decoders. To step further, we also introduce a discriminator to better align the learned representations for different graphs via adversarial training. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of the proposed approach.