Alleviating the over-smoothing of graph neural computing by a data augmentation strategy with entropy preservation

Highlights：

• We provide a new graph entropy design to measure the smoothness of graph feature manifold and conclude that motif-based information structures determine this graph entropy.

• We propose a novel graph data augmentation strategy that protects not only the integrity of topological structure but also the integrity of motif-based information structures. Our strategy shows advantages in maintaining more graph entropy compared with other methods.

• We demonstrate our model on different datasets, and experiments show that our work is superior in real-world node classification tasks.

• Our approach significantly enhances the robustness of GCN and could alleviate the over-smoothing phenomenon to a certain extend.