Multi-Scale Variational Graph AutoEncoder for Graph Disentanglement Learning

Graph Convolutional Neural Networks are serving as a common network to learn complex relationships and extract deep features for downstream tasks. The existing graph-based autoencoders take an encoder with fixed adjacency matrix to learn latent representations in the real graph and adopt a decoder to reconstruct the adjacency matrix for self-supervised learning. However, the relations are heterogeneous in the real world, but entangled together because of many latent factors. In this paper, we propose a novel graph framework, termed multi-scale variational graph autoencoder for graph disentanglement learning, which disentangles the input graph into several factor graphs and learns latent representations. Then, we present a decoder to infer mixed relations from latent representations. In addition, we introduce β-VAE framework to maintain diversity of latent representations. Generally, different metrics demonstrate our model’s advantaged performance on both downstream tasks and graph-level disentanglement tasks. Our model outperforms the state-of-the-art methods, especially on the ZINC dataset and other datasets, for example those related to bioinformatics graphs, which are featured by practical node representations.