Abstract:Graph neural network(GNN) models have been widely used in graph-structured data which can be seen everywhere but difficult to deal with. Link prediction is a kind of typical graph learning task which has been widely used in many applications such as knowledge graph completion, social recommendation, etc. Most existing link prediction models need preprocessing to preserve properties of graph such as asymmetric transitivity, node proximity, etc., which make models inconvenient to use. In this paper, we focus on link prediction in directed graphs and propose an end-to-end model called directed graph variational auto-encoder(DGVAE). DGVAE uses variational auto-encoder(VAE) as the whole framework and uses a two-layer GNN model FDGCN as the encoder. We design a brand new decoder for asymmetric output. Experiments are conducted on several different datasets and existing state-of-the-art models. Results show that DGVAE has the best performance in most cases.

DGVAE:An End-to-end Model for Link Prediction in Directed Graphs