Multi-level disentanglement graph neural network

Real-world graphs are generally generated from highly entangled latent factors. However, existing deep learning methods for graph-structured data often ignore such entanglement and simply denote the heterogeneous relations between entities as binary edges. In this paper, we propose a novel Multi-level Disentanglement Graph Neural Network (MD-GNN), a unified framework that simultaneously implements edge-level, attribute-level, and node-level disentanglement in an end-to-end manner. MD-GNN takes the original graph structure and node attributes as input and outputs multiple disentangled relation graphs and disentangled node representations. Specifically, MD-GNN first disentangles the original graph structure into multiple relation graphs, each of which corresponds to a latent and disentangled relation among entities. The input node attributes are then propagated in the corresponding relation graph through a multi-hop diffusion mechanism to capture long-range dependencies between entities, and finally the disentangled node representations are obtained through information aggregation and merging. Extensive experiments on synthetic and real-world datasets have shown qualitatively and quantitatively that MD-GNN yields truly encouraging results in terms of disentanglement and also serves well as a general GNN framework for downstream tasks. Code has been made available at: https://github.com/LirongWu/MD-GNN.