Nguyen Thanh Toan¹, Phan Thanh Cong² and Quan Thanh Tho¹ , ¹Ho Chi Minh City University of Technology, Vietnam and ²Griffith University, Australia

Network alignment, the task of seeking the hidden underlying correspondence between nodes across networks, has become increasingly studied as an important task to multiple network analysis. A few of the many recent applications of network alignment include protein network alignment, social network reconciliation, and computer vision. However, traditional methods which are based on matrix factorization directly work on networks themselves rather than exploit their intrinsic structural consistency, and thus their performance is sensitive to structural variations of networks. Recently, many supervised approaches which leverage latent representation have been proposed. Although they can handle large-scale datasets, most of them rely on a large number of parallel anchor links which are unavailable or expensive to obtain for many domains. Therefore, in this paper, we propose the WENA Framework, a representation learningbased network alignment, in which we study how to design weakly-supervised methods to align large-scale networks with a limit of ground truth available. Empirical results show that, with only two anchor links, WENA significantly outperforms existing unsupervised aligners and even outperforms state-of-the-art supervised methods that use richer resources in terms of both noise robustness and accuracy.

Network embedding, Graph mining, Network alignment, Graph matching, Knowledge representation