We provide code to reproduce results shown in the paper "A Comparative Study on Robust Graph Neural Networks to Structural Noises". The paper can be viewed here.

Our original paper compares the model performance of eight mainstream robust GNNs under consistent noise settings. The main contribution is the design of three structural noises under different granularity: local, community, and global. The noise generation methods can be found in the noise.py file.

For simplicity, we only give an example of how to use the noise generation module based on the GraphSAGE. The basic workflow is: perturbing a clean graph by calling the implemented noise functions; afterward, feeding the poisoned graph to a robust model.