Moscat: Mixture of Scope Experts at Test

Code repository for paper: Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants

⚙️ Setup

Ensure your environment meets the following dependencies (newer versions might also work):

• Python 3.11.4
• PyTorch 2.0.1
• torch_geometric 2.4.0
• torch-scatter 2.1.2
• torch-sparse 0.6.18
• hydra-core 1.3.2
• hydra-colorlog 1.2.0
• torchmetrics 0.11.4
• class_resolver 0.4.3

🚀 Launching Moscat

All configuration files are available in the conf directory to help reproduce the results reported in Table 1 (Main Results) and Table 7 (Leaderboard Comparison).

Note: You can adjust the logging verbosity by setting log_level=INFO (default) for more detailed logs or log_level=CRITICAL for minimal logging output.

1. Train GNN Scope Experts (Scope 0 to 6)

• 0-hop MLP Scope Expert:

  python run_gnn.py -m +exp_gnn=penn94_mlp log_logit=true  

• 1-6 hop GNN Scope Experts:

  python run_gnn.py -m +exp_gnn=penn94_gcn model.conv_layers=1,2,3,4,5,6 log_logit=true

2. Train the Moscat Gating Model

python run_moscat.py +exp_moscat=penn94_gcn