GRAFT (Gradient-Free Editing for GNNs)

GRAFT contains code and shell drivers for gradient-free editing experiments on graph neural networks.

Overview

• run_*.sh scripts live under ablations/, editing_pipelines/, and seed-gnn/ and are the usual entrypoints for batch jobs.
• Path defaults: Repo-root paths.py and paths.sh set PATH_TO_GRAFT (default: the repository containing those files, typically /home/model_editing/GRAFT) and PATH_TO_DATA (default /home/model_editing/data). Shell drivers source paths.sh; override locations with env vars (DATASET_DIR, PRETRAIN_DIR, etc. still work as documented in each script).
• Dataset preparation: Import raw graphs dataset-by-dataset from the canonical repos, preprocess to match this codebase’s CSV layout under DATASET_DIR (see paths.sh), and respect each publisher’s license.
  • GADBench — anomaly / fraud benchmarks used here include YelpChi (mapped as yelp) and T‑Finance (mapped as tfinance); use its datasets tree and preprocessing notebook as in that README.
  • GraphLand — regression targets such as artnet-views and twitch-views (run_node_regression.sh): follow data plus dataset.py / PyGDataset in that repo.
  • Pokec — take the release from GUIDE (KDD 2022; Group Equality Informed Individual Fairness in GNNs): files live under dataset (pokec is distributed zipped there—unzip before use), then convert to this repo’s CSV layout under DATASET_DIR.
  • Bail — take the fairness benchmark graphs from NIFTY (Towards a Unified Framework for Fair and Stable Graph Representation Learning, arXiv:2102.13186): data live under dataset (that repo README notes large edge bundles may be zipped), then preprocess to match DATASET_DIR locally.
• More detail on layout and configs: editing_pipelines/README.md, seed-gnn/README.md.

Environment

The working environment is captured as conda env gnn_edit:

1. Full conda solve (recommended) — reproduces Python, CUDA-related conda packages, and pip extras:

 conda env create -f environment.yml
 conda activate gnn_edit

2. Pip-only pin list — requirements.txt lists PyPI packages as pinned in that env. Some lines (PyTorch +cu118, pyg-lib / torch-scatter, etc.) normally need the official PyTorch / PyG wheel indexes; if pip install -r requirements.txt fails on those, install them the way PyTorch and PyG document for your CUDA version, or stick to environment.yml.

environment.yml was produced with conda env export -n gnn_edit --no-builds and without the prefix: field so paths stay machine-agnostic.

Experiment scripts (run_*.sh)

Editing pipelines (editing_pipelines/)

Hub

• run_edit.sh — Wraps run_edit.py: method, dataset, model, seed, target counts, max steps, strategies, and data/pretrain/output paths. Reads optional hyperparameters from the environment (LAMBDA_REG, TOP_FRACTION, NUM_LAYERS, GAMMA_RETAIN, FT_EPOCHS, FT_LR, PR_ALPHA, RANK_MIX_TAU). Sets PYTHONPATH via paths.sh.

Parameter sweeps (loop and call run_edit.sh)

• run_editing_suite.sh — Large least-squares grid: λ, top fraction, depth, many LS_STRATEGY modes. Uses run_edit.sh with --debug.
• run_seed_gnn.sh — Seed-GNN baseline over datasets, models, seeds, top fractions, depth (NUM_LAYERS / LAYERS_OVERRIDE).
• run_egnn.sh — EGNN baseline; same idea, optional layers_<n>/top_<frac> paths when layers are overridden.
• run_finetune.sh — Finetune baseline over models, seeds, top fraction, FT_EPOCHS / FT_LR, and depth.

Ablations (ablations/)

• run_feature_ablation_all_checkpoints.sh — Finds metrics_*.json under OUTPUT_ROOT for pokec / bail / yelp, GCN_MLP / GIN_MLP / Polynormer, and methods leastsquares / finetune / egnn / seed_gnn. Delegates to launch_feature_ablation_from_metrics.py → feature_ablation_forward.py. Supports FEATURE_ABLATION_FILTER_*, DRY_RUN=1, and FEATURE_ABLATION_SKIP_EXISTING.

Pretraining (seed-gnn/)

• run.sh — Main vanilla pretraining loop: configure the datasets, models, seeds, and num_layers arrays plus optional feature-ablation block at the top of the script, then it calls scripts/pretrain/seed_gnn/<dataset>.sh per combination (two path arguments: output root and dataset dir appear as ~/data/seed_gnn_data / ~/data/seed_gnn_data/dataset in the defaults—adjust to match your paths.sh layout). Accepts --architectures, --gat_optimized, --neighbor_batch_size, and --neighbor_num_neighbors forwarded to each pretrain invocation.
• run_node_regression.sh — Pretrain on node-regression sets (e.g. twitch-views, artnet-views); sweeps gcn/sage/gin/gat, seeds, depth via scripts/pretrain/seed_gnn/.
• run_feature_drop_pretrain_suite.sh — Pretrain with one sensitive column dropped per dataset (default pokec/bail/yelp/tfinance), aligned with editing suite attribute names; outputs under edit_ckpts_feature_ablated/.../no_<feature>/. Arguments after -- forward to main.py.
• run_feature_drop_pretrain_regression.sh — Same pattern for regression graphs and their sensitive columns.