Directed Graph Auto-Encoders

We introduce a new class of auto-encoders for directed graphs, motivated by direct extension the Weisfeiler-Leman algorithm to pairs node labels. The proposed model learns interpretable latent representations nodes and uses parameterized graph convolutional network (GCN) layers its encoder an asymmetric inner product decoder. Parameters in control weighting exchanged between neighboring nodes. demonstrate ability learn meaningful embeddings achieve superior performance on link prediction task several popular datasets.