Link Prediction via Factorization Machines

Abstract

Link prediction is the problem of predicting the existence of edges in a network. The link prediction problem is a fundamental research problem in graph mining and has numerous applications in social networks, bioinformatics, e-commerce, etc. A major challenge of link prediction problems is handling the fact that real-world networks are becoming extremely large. The large network size leads to huge sparsity in the network’s adjacency matrix which most existing link prediction methods (such as matrix factorization) rely on. Moreover, when networks become very large, there exists a non-trivial link imbalance problem where the numbers of known present and known absent links are significantly different. Such sparsity and imbalance issues significantly impact and decrease the performance of existing link prediction methods. To address these challenges, in this research we propose a Balanced Factorization Machine (BFM) which performs link predictions on very sparse network via learning interactions among nodes and edges of the network in a supervised learning setting. Through extensive experiments on real-world network data sets, we show that our BFM method significantly outperforms other existing link prediction methods.