GnnXemplar: Exemplars to Explanations - Natural Language Rules for Global GNN Interpretability

This repository contains the official implementation of GnnXemplar, accepted as an oral presentation at NeurIPS 2025.

• Paper: https://neurips.cc/virtual/2025/loc/san-diego/poster/116901
• Oral presentation: https://neurips.cc/virtual/2025/loc/san-diego/session/122572

To cite our work, use:

@inproceedings{armgaan2025gnnxemplar,
    author = {Armgaan, Burouj and Jain, Eshan and Pandey, Harsh and Chandran, Mahesh and Ranu, Sayan},
    title = {{GnnXemplar}: Exemplars to Explanations - Natural Language Rules for Global {GNN} Interpretability},
    booktitle = {Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS)},
    year = {2025}
}

Environment

conda clean -a
pip cache purge
conda create -n gnnxemplar python=3.10
conda activate gnnxemplar

# torch, torch_geometric
# We've used a cpu-only version. Feel free to use a gpu-version if you like.
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 cpuonly -c pytorch
conda install pyg=2.5.2 -c pyg

# general purpose
conda install scikit-learn=1.4.2 networkx=3.1

# ogb
pip install ogb==1.3.6

# gemini
pip install google-generativeai==0.8.3

Prompts and Models

Prompts and trained models can be found in data/ folder which has the following directory structure:

data/
├── amazonratings
│   ├── dataset_config.json
│   └── model.pt
├── arxiv
│   ├── dataset_config.json
│   └── model.pt
├── bashapes
│   ├── dataset_config.json
│   └── model.pt
├── Citeseer
│   ├── dataset_config.json
│   └── model.pt
├── minesweeper
│   ├── dataset_config.json
│   └── model.pt
├── questions
│   ├── dataset_config.json
│   └── model.pt
├── TAGCora
│   ├── categories.csv
│   ├── cora.pt
│   ├── dataset_config.json
│   └── state_dict.pt
└── WikiCS
    ├── dataset_config.json
    └── state_dict.pt

Running the code

• Update your gemini API Key in src/llm_self_refine.py
• Run all commands from the repo's home directory.

# This finds the exemplars for each class and store them in the data/ directory.
python src/find_exemplars.py -h
python src/find_exemplars.py ...

# This finds the global explanations and stores them in the data/ directory.
python src/llm_self_refine.py -h
python src/llm_self_refine.py ...

Example run on TAGCora with the default parameters

python src/find_exemplars.py -d TAGCora 
python src/llm_self_refine.py -d TAGCora

Output explanations

After running the pipeline on a dataset, you'll find:

• Signatures in data/<dataset>/cluster_summaries.txt
• Human-readable global explanations in data/<dataset>/combined_explanations.txt
• Metrics in data/<dataset>/metrics.txt