Adaptive Multi-Channel Deep Graph Neural Networks

Abstract

Graph neural networks (GNNs) have shown significant success in graph representation learning. However, the performance of existing GNNs degrades seriously when their layers deepen due to the over-smoothing issue. The node embedding incline converges to a certain value when GNNs repeat, aggregating the representations of the receptive field. The main reason for over-smoothing is that the receptive field of each node tends to be similar as the layers increase, which leads to different nodes aggregating similar information. To solve this problem, we propose an adaptive multi-channel deep graph neural network (AMD-GNN) to adaptively and symmetrically aggregate information from the deep receptive field. The proposed model ensures that the receptive field of each node in the deep layer is different so that the node representations are distinguishable. The experimental results demonstrate that AMD-GNN achieves state-of-the-art performance on node classification tasks with deep models.