LON-GNN: Spectral GNNs with Learnable Orthonormal Basis

Source code for "LON-GNN: Spectral GNNs with Learnable Orthonormal Basis".

Requirements

• Python 3.7
• numpy 1.21
• pytorch 1.12
• pyg 1.7.1
• optuna 3.0.5

Experiments on Fitting Images

Codes on fitting images can be found in LON-GNN/LearningFilters/. To reproduce the results of LON-GNN, run the following commands:

cd LON-GNN/LearningFilters/
./test_LONGNN.sh

The detailed parameters can also be found in LON-GNN/LearningFilters/test_LONGNN.sh.

Experiments on Real-World Datasets

Reproduce LON-GNN

To reproduce the results of LON-GNN on real-world datasets, run the following commands:

cd LON-GNN/
./test_LONGNN.sh

The detailed parameters can also be found in LON-GNN/test_LONGNN.sh.

Users can also utilize optuna to search the best parameters:

cd LON-GNN/
./param_search.sh ${dataset} 0

where ${dataset} can be selected from cora, pubmed, citeseer, computers, photo, actor, chameleon, squirrel, cornell and texas.

Ablation

To reproduce Jacobi+Orthnorm in ablation analysis, run the following commands:

cd LON-GNN/
./test_Jacobi+Orthnorm.sh

To reproduce Jacobi+Learnable in ablation analysis, run the following commands:

cd Jacobi+Learnable
./test_Jacobi+Learnable.sh