Netmerger: Predicting Cross-Network Links In Merged Heterogeneous Networks

Previous work on link prediction focuses on single network settings or transferring knowledge between two networks for predicting intra-network links. However, many real-world applications involve multiple networks, and we need to predict cross-network links in these networks instead. For example, when mergers or acquisitions happen between two companies, the two business networks need to be merged into one. We may want to promote cross-network links, such as connections between the users in PayPal with those in x.com, to speed up the merger and integration of the two business networks. Considering recommending cross-network links can significantly speed up the business integration, in this paper, we study the problem of cross-network link prediction in merged heterogeneous networks. The goal is to predict potential links between multiple kinds of nodes across the two merged networks. It is also highly challenging because little or no information is previously known about the cross-network links, i.e., a cold-start problem. We proposed an approach, called NetMerger, which first learns network embeddings of the merged networks based upon their network structures and then makes cross-network link predictions. We compared our method with baselines on two real-world datasets. The results demonstrated the effectiveness of our proposal.