Revisiting Over-smoothing in Deep GCNs
This work addresses performance issues in deep GCNs for graph-based learning tasks, offering a novel perspective and practical improvement, though it appears incremental as it builds on existing GCN architectures.
The paper challenges the assumption that oversmoothing is the main cause of performance drop in deep graph convolutional networks (GCNs), proposing that deep GCNs can learn anti-oversmoothing during training, and introduces a cheap trick to improve GCN training, verified on three citation networks.
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.