xwangxshi/fspecgnn-heterophily-checkpoints

FSpecGNN Heterophily Checkpoints

This repository contains 10-run checkpoints for the FSpecGNN heterophily experiments on the small and large benchmark groups.

Each run directory follows the common PyTorch checkpoint layout:

checkpoints/{small,large}/{Dataset}/{Model}/run_XX_seed_XXXXXXXXXX/
├── pytorch_model.bin
├── config.json
├── metrics.json
├── training_args.json
└── split_masks.pt

The manifest.jsonl file contains one sanitized record per checkpoint with relative paths only. The metrics/ directory contains the per-run summary and aggregate results.

Results

group	dataset	model	metric	mean ± ci95
large	Minesweeper	Bern_GAT	roc_auc	84.6384 ± 2.1314
large	Minesweeper	ChebII_GAT	roc_auc	85.3813 ± 1.8125
large	Minesweeper	Cheb_GAT	roc_auc	89.0344 ± 0.5564
large	Questions	Bern_GAT	roc_auc	74.9557 ± 0.8790
large	Questions	ChebII_GAT	roc_auc	75.5476 ± 0.5814
large	Questions	Cheb_GAT	roc_auc	76.6198 ± 0.5161
large	Roman_empire	Bern_GAT	accuracy	57.0567 ± 1.0361
large	Roman_empire	ChebII_GAT	accuracy	57.2322 ± 1.1752
large	Roman_empire	Cheb_GAT	accuracy	56.5162 ± 1.4473
large	Tolokers	Bern_GAT	roc_auc	77.1383 ± 0.9651
large	Tolokers	ChebII_GAT	roc_auc	77.4562 ± 0.9763
large	Tolokers	Cheb_GAT	roc_auc	78.2609 ± 0.6133
small	Chameleon	Bern_GAT	accuracy	37.6769 ± 2.4065
small	Chameleon	ChebII_GAT	accuracy	36.3325 ± 3.1843
small	Chameleon	Cheb_GAT	accuracy	36.4269 ± 1.5094
small	Squirrel	Bern_GAT	accuracy	34.7017 ± 4.3088
small	Squirrel	ChebII_GAT	accuracy	35.9186 ± 2.8930
small	Squirrel	Cheb_GAT	accuracy	40.2083 ± 0.6964
small	Texas	Bern_GAT	accuracy	62.5434 ± 4.5694
small	Texas	ChebII_GAT	accuracy	60.9249 ± 3.4118
small	Texas	Cheb_GAT	accuracy	54.7977 ± 4.5665
small	Wisconsin	Bern_GAT	accuracy	54.5000 ± 5.7510
small	Wisconsin	ChebII_GAT	accuracy	53.5000 ± 4.3333
small	Wisconsin	Cheb_GAT	accuracy	45.3750 ± 7.5062

For binary large datasets (Minesweeper, Tolokers, and Questions), the recorded metric is ROC-AUC. Other rows use accuracy.

Loading A Checkpoint

import json
import torch

ckpt_dir = 'checkpoints/small/Wisconsin/ChebII_GAT/run_01_seed_4198936517'
config = json.load(open(f'{ckpt_dir}/config.json'))
state_dict = torch.load(f'{ckpt_dir}/pytorch_model.bin', map_location='cpu')

Instantiate the matching model architecture from the FSpecGNN codebase using the fields in config.json, then load state_dict.