Network Alignment by Representation Learning on Structure and Attribute

Network alignment is the task of recognizing similar network nodes across different networks, which has many applications in various domains. As traditional network alignment methods based on matrix factorization do not scale to large graphs, a variety of representation learning based approaches has been proposed recently. However, these techniques tend to focus on topology consistency between two networks while ignoring other valuable information (e.g. network nodes attribute), which makes them susceptible to structural changes. To alleviate this problem, we propose RAN, a representation-based network alignment model that couples both structure and node attribute information. Our framework first constructs multi-layer networks to represent topology and node attribute information, then computes the alignment result by learning the node embeddings for source and target network. The experimental results show that our method is able to outperform other techniques significantly even on large datasets.