Link prediction approach combined graph neural network with capsule network

Graph Neural Networks (GNNs, in short) are a powerful computational tool to jointly learn graph structure and node/edge features. They achieved an unprecedented accuracy in the link prediction problem, namely the task of predicting if two nodes are likely to be tied by an edge in the near future. However, GNNs capture node attributes as scalars and such a representation can be non-optimal. In this paper we propose a link prediction approach which combines GNN with Capsule Networks (CapsNet), a powerful deep learning framework that obtained high quality representations when applied to image processing. Our approach first applies a GNN to generate node embeddings. Then, it makes use of a conversion block with the purpose of transforming node embeddings into a node pair feature map (called edge feature map). Because of such a transformation, the link prediction problem is equivalent to a graph classification problem. Finally, our approach applies CapsNets to learn the feature representation of the edge feature map, so that the attributes of the node pairs are captured from different aspects. We evaluate the feasibility and effectiveness of our approach (called Graph Conversion Capsule Link, GCCL in short) on six networks without node attributes and three networks with node attribute. Experimental results show that in both conditions (with and without node attributes) our approach is significantly more accurate than the competitor methods, with an average accuracy improvement of about 20%.