Abstract
Recent advances in representation learning produce node embeddings that may be used successfully in many downstream tasks (e.g., link prediction), but do not generally extend beyond a single network. Motivated by the prevalence of multi-network problems, such as graph alignment, similarity, and transfer learning, we introduce an elegant and principled node embedding formulation, Cross-network Matrix Factorization or xNetMF, to effectively and efficiently learn embeddings from similarities based on the nodes' structural and attribute identities (if available), which are comparable across networks. Leveraging the Nystr\"om method for low-rank approximation, we achieve significant reduction in space and time compared to the na\"ive approach of directly factorizing the node similarity matrix, with xNetMF running up to 30x faster than existing representation learning methods. We study its effectiveness on the task of network alignment, and introduce REGAL (REpresentation-based Graph ALignment), the first method to use node representation learning for this task. Extensive experiments on a variety of challenging real-world networks with and without attributes show that REGAL outperforms existing network alignment methods by 20 to 30% in accuracy on average, uses attributes up to 40x faster, and scales to networks with millions of nodes each.

This publication has 11 references indexed in Scilit: