Structural representation learning for network alignment with self-supervised anchor links

Abstract

Network alignment, the problem of identifying similar nodes across networks, is an emerging research topic due to its ubiquitous applications in many data domains such as social-network reconciliation and protein-network analysis. While traditional alignment methods struggle to scale to large graphs, the state-of-the-art representation-based methods often rely on pre-defined anchor links, which are unavailable or expensive to compute in many applications. In this paper, we propose NAWAL, a novel, end-to-end unsupervised embedding-based network alignment framework emphasizing on structural information. The model first embeds network nodes into a low-dimension space where the structural neighborhoodship on original network is captured by the distance on the space. As the space for the input networks are learnt independently, we further leverage a generative adversarial deep neural network to reconcile the spaces without relying on hand-crafted features or domain-specific supervision. The empirical results on three real-world datasets show that NAWAL significantly outperforms state-of-the-art baselines, by over 13% of accuracy against unsupervised methods and on par or better than supervised methods. Our technique also demonstrate the robustness against adversarial conditions, such as structural noises and graph size imbalance.