Reduce Over-smoothing in Label Non-Uniformity Learning Based on Graph Convolution Network

In the semi-supervised training of Graph Convolution Network (GCN), the output signals can be used to reveal the graph structure to some extent. By exploiting this property, the label non-uniformity training procedure attempts to extract high confidence nodes and drop the low important edges to increase the learning capacity of the base model. However, this approach leads to the increasing the over-smoothing issue in the learning process. Our works begin to examine the over-smoothing degrees in multi-step training with label non-uniformity weights through the mean average distance measure. Based on this observation, we propose the usage of residual connections to improve the base GCN. The resulting model can reduce the over-smoothing effect, and thus improve the learning capacity of the model. We evaluate our approach on 6 public datasets and the results show that our proposed model can improve the results of the GCN architecture and other baselines with a wide margin.