Rethinking Spectral Graph Neural Networks with Spatially Adaptive Filtering

This repository contains a PyTorch implementation of the work "Rethinking Spectral Graph Neural Networks with Spatially Adaptive Filtering".

Environment Settings

• pytorch 1.8.0
• numpy 1.19.5
• torch-geometric 1.7.2
• scipy 1.5.3
• seaborn 0.11.2
• sklearn 0.24.2
• pickle 4.0
• optuna (for hyper-parameter search)

Datasets

We provide one dataset in the folder './data'. Other datasets can be downloaded through the links provided in the Appendix.

Code Structure

1. Preprocessing

Run the script below to preprocess the datasets in ./data/raw/ (the preprocessed data will be saved in ./data/processed and ./data/eigen_dcp).

python preprocessing.py

2. Generating random data splits

Generate the random node-classification splits with gen_splits.py. Following the paper, two supervision modes are produced for each dataset:

• full-supervised — 60% / 20% / 20% (train / val / test), saved as *_denseSplits.npy
• semi-supervised — 2.5% / 2.5% / 95% (train / val / test), saved as *_sparseSplits.npy

For each mode, nb_split (default 10) independent splits are created. The training nodes are sampled class-balanced, the validation nodes are drawn at random from the rest, and all remaining nodes form the test set. The files are written to ./data/random_splits/ in the format expected by main.py.

python gen_splits.py --dataset squirrel --nb_split 10

(Run python preprocessing.py first, as the labels are read from the processed data.)

3. Searching hyper-parameters with Optuna

Use optuna_search.py to tune the model. The search space follows Sec. VI-A of the paper: learning rate, weight decay (L2), dropout, nonlocal aggregation steps L, scaling tau, update rate eta, and (for SAF_eps) the sparsification threshold eps. Each trial is scored by the mean validation accuracy over the first --n_eval_splits random splits, and the best hyper-parameters are dumped to a JSON file (best_params.json by default).

python optuna_search.py --model SAF     --dataset squirrel --sl_mode full --n_trials 100
python optuna_search.py --model SAF     --dataset squirrel --sl_mode semi --n_trials 100