Approaching feature matrix: To solve two issues in link prediction

Abstract

With the increase of graph-structure data, link prediction has become an active research topic. Recent progress in link prediction involves learning graph embeddings through graph neural networks. Graph autoencoders(GAE), variational graph autoencoders(VGAE), and their variants such as graph normalized autoencoders(GNAE) try to learn graph embeddings in an unsupervised setting and show promising results in link prediction. However, they suffer from degraded performance with two issues; they tend to assign close to-zero embedding on nodes with zero degree (i.e., isolated node problem) and the issue caused by random negative sampling has not been addressed in link prediction experiments. As the two issues affect only links and not feature information, we propose a generic learning method based on the property of feature information that alleviates both issues and outperforms strong baselines across benchmark datasets.