Spectre

Mattie Ji | Amauri H. Souza | Vikas Garg

This is the official repo for the paper Graph persistence goes spectral (NeurIPS 2025).

Installing routines to compute SpectRe (FastSpectre) diagrams

This repo is highly based on https://github.com/Aalto-QuML/RePHINE. Please, see the base repo for installing basic dependencies.

cd torch_ph
python setup_spectre.py install 
python setup_spectre_scheduling.py install

Experiments on expressivity

To reproduce the experiments on Cayley graphs, e.g., cayley-24, you must first run

python datasets/create_cayley_data.py --dataset minCayleyGraphs24Vertices

Then, see the python notebook Experiments - Cayley.ipynb to obtain the results.

For BREC datasets, run the python notebook Experiments - BREC.ipynb.

Experiments using real data

For the main experiments, we run the main.py with the arguments in cli_main.py. For instance, to run FastSpectre on NCI109 combined with a GCN model, we run:

python main.py --dataset NCI109  --diagram_type fast_spectre --gnn gcn --max_epochs 200 --out_dim_eigen_deep_set 32 --num_filtrations 2  

Citation

@inproceedings{spectre,
title={Graph Persistence goes Spectral},
author={Mattie Ji and Amauri H. Souza and Vikas K. Garg},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
year={2025},
url={https://openreview.net/forum?id=wU8IKGLpbi}
}