Skip to content
bckrlabPublic

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

Repository files navigation

bk-buster

Code for the synthetic node-classification benchmark from:

K. Coşkun, F. Blanke, V. S. Cangalovic, B. C. Hiller, A. Mirzaei, T. Siegl, S. Lüdtke, M. Becker. Informed, but Not Always Improved: Challenging the Benefit of Background Knowledge in GNNs. IEEE Transactions on Computational Biology and Bioinformatics.

This benchmark (Section V.D / Fig. 11 in the paper) is a controlled, synthetic node-classification task used as a positive control: node features are drawn from overlapping distributions, and a background-knowledge (BK) graph connects nodes of the same class. It lets us test how BK-informed GNNs (GCN/GAT) respond to graph perturbations (edge removal, edge addition, node isolation, detach-and-rewire) compared to uninformed baselines (LogReg, SVM, MLP), in a setting where the BK graph is guaranteed to be informative by construction.

Installation

The dependencies (PyTorch, PyG, RAPIDS, etc.) are pinned in dev/env.yml.

Option A — Docker (recommended, includes GPU support):

cd dev
docker compose up -d --build
docker compose exec app bash

This builds an image with a dev micromamba environment and installs the package into it in editable mode (see the command in dev/compose.yml).

Option B — conda/micromamba directly:

micromamba env create -f dev/env.yml -n dev
micromamba activate dev
pip install -e .

Notebooks

All notebooks live under exps/syn_node/.

  • data_generation.ipynb — builds the synthetic node-classification dataset (overlapping node-feature distributions, same-class BK graph) and visualizes it, along with each perturbation type (edge removal/addition, node isolation, detach-and-rewire) applied to an example graph.
  • binary_test.ipynb — runs the actual benchmark: trains GCN/GAT (informed) against LogReg/SVM/MLP (uninformed baselines) over 10 runs, for each perturbation type at increasing severity, and reproduces Fig. 11 of the paper.

Running binary_test.ipynb writes results to exps/syn_node/syn_node_{rem_edges,add_edges, iso_nodes,dr_nodes}.csv and the corresponding .pdf figures. These outputs are gitignored — rerun the notebook to regenerate them.

Package layout

License

MIT — see LICENSE.txt.


This project was set up using PyScaffold 4.6.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages