Link prediction based on graph convolutional networks utilizing normalization and a communicability kernel

Abstract

Link prediction is to predict whether there is a link between two nodes in a network. Graph Convolutional Networks have outperformed existing works in many graph analysis tasks. Most previous methods propagate neighbor nodes without normalization for graph convolutions. This makes optimization during the learning process tends to be more affected by the norms of embedding vectors while treat directions of embedding vectors lightly. In this paper, we devised a novel graph convolutional network to which the normalization technique was applied to prevent the result from being biased against the size of the norm of the embedding vector during link prediction. To leverage the information of higher-order relationships for link prediction. We also propose a new communicability filter that can effectively reflect higher order information. Our proposed higher-order graph convolutional networks can perform link prediction more accurately. Through extensive experiments with various relational datasets, we show that the proposed methods outperforms the state-of-the-art link prediction methods on heterogeneous graphs as well as homogeneous graphs.