Impact-of-heterophilic-designs

The relevant codes for "What Contributes More to the Robustness of Heterophilic Graph Neural Networks?", [Under Review].

• Requirements

  • Our codes are built based on the project GraphWar (Now GreatX) , so you may need to install PyTorch, PyTorch Geometric, and GraphWar first. Please see the 'requirements.txt' for details.

• Code illustrations

  • examples/attack/targeted: Main demo folder

    • for_real_gnn_evaluate.py: main codes for conducting attacks on realistic GNNs, such as H2GCN and UGCN.
    • sim_heterophilic_attack.py: main codes for conducting attacks on baseline GCN models integrating with different heterophilic designs.

  • graphwar/heter_gnn: Main algorithm folder

    • basic_gcn.py
    • h2gcn.py
    • ugcn.py

  • heterophily_dataseets_matlab: Folder of the corresponding heterophilic graph data.

• Run the demo

  # baseline GCN model integrating with different heterophilic designs
  python sim_heterophilic_attack.py
  
  # realistic GNN model
  python for_real_gnn_evaluate.py