Noisy-Graph-Active-Learning

Official implementation of GALClean+ (Graph Active Learning and Cleaning) from the paper "Active Learning for Graphs with Noisy Structures".

Environment

Tested with Python 3.10 and PyTorch 2.5:

conda create -n galclean python=3.10 pip
conda activate galclean
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install numpy scipy scikit-learn pandas matplotlib seaborn tqdm networkx pyyaml ogb
pip install torch_geometric==2.6.1
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.5.0+cu121.html

Quick Start

Run GALClean+ on Cora with default settings:

python train.py --data Cora --noise_level 0.0 --rand_seed 0 --init_split 0

Outputs are written to outputs/ as:

• outputs/<run_id>/metrics.json
• outputs/<run_id>/selected_indices.json

metrics.json records the paper-aligned settings used for the run.

Defaults (paper-aligned in this repo):

• Planetoid split: public_fixed
• Random noise sampling: any
• Edge filtering: legacy

Splits: Planetoid (Cora, CiteSeer, PubMed) uses the legacy fixed split (public_fixed). Amazon/Coauthor use random.

Budget: Default is 2 initial nodes/class + 8 selected nodes/class = 10 total/class.

Experiments

Random Edge-Adding Attack:

python train.py --data Cora --noise_level 0.6

Unsupervised Adversarial Attack (CLGA):

python train.py --data Cora --attack clga --attack_percent 20

This repo ships the paper CLGA attacked graphs under data/attacks/clga/: Cora/CiteSeer/PubMed/Amazon-Photo/Coauthor-CS at 5/10/15/20%.

Sweeps + plots

Random noise sweep (defaults to noise levels 0.0..1.0 step 0.2):

python scripts/run_sweep.py --mode random --datasets Cora CiteSeer PubMed --seeds 0 --splits 0
python scripts/collect_results.py
python scripts/plot_results.py --scale_percent

CLGA sweep:

python scripts/run_sweep.py --mode clga --datasets Cora CiteSeer PubMed Amazon-Photo Coauthor-CS --seeds 0 --splits 0
python scripts/collect_results.py
python scripts/plot_results.py --scale_percent

Citation

@article{chi2024active,
  title={Active Learning for Graphs with Noisy Structures},
  author={Chi, Hongliang and Qi, Cong and Wang, Suhang and Ma, Yao},
  journal={arXiv preprint arXiv:2402.02321},
  year={2024}
}