Revisiting Over-smoothing in Deep GCNs

TL;DR

This paper challenges the traditional view of oversmoothing in deep GCNs by showing that they can learn to counteract oversmoothing during training, and proposes a simple method to enhance training effectiveness.

Contribution

It offers a new perspective that deep GCNs can learn anti-oversmoothing, and introduces an effective training trick based on this insight.

Findings

Deep GCNs learn anti-oversmoothing during training.

A simple trick improves GCN training performance.

Analysis of neighborhood aggregation in GCNs.

Abstract

Oversmoothing has been assumed to be the major cause of performance drop in deep graph convolutional networks (GCNs). In this paper, we propose a new view that deep GCNs can actually learn to anti-oversmooth during training. This work interprets a standard GCN architecture as layerwise integration of a Multi-layer Perceptron (MLP) and graph regularization. We analyze and conclude that before training, the final representation of a deep GCN does over-smooth, however, it learns anti-oversmoothing during training. Based on the conclusion, the paper further designs a cheap but effective trick to improve GCN training. We verify our conclusions and evaluate the trick on three citation networks and further provide insights on neighborhood aggregation in GCNs.