Joint Denoising and Rewiring

This repository contains the code for the ICLR2025 [Paper] "Joint Graph Rewiring and Feature Denoising via Spectral Resonance".

The GNN code is based on the ICLR2021 paper Adaptive Universal Generalized PageRank Graph Neural Network [Paper] [Code].

Baseline Methods

Diffusion Improves Graph Learning (DIGL) [Paper] [Code]

First-order spectral rewiring (FoSR) [Paper] [Code]

Batch Ollivier-Ricci Flow (BORF) [Paper] [Code]

Approximate Message Passing - Belief Propagation (AMP-BP) [Paper] [Code]

DIGL, FoSR and BORF can be run directly from this repository. For AMP-BP, please refer to the original repository.

Requirements:

Tested with Python 3.10.14 and PyTorch 2.0.1 (Cuda 11.8).

pytorch
pytorch-geometric
numpy scipy matplotlib pyyaml

For FoSR and BORF baseline:

numba pandas networkx GraphRicciCurvature scikit-learn

Optional (if not used, use flag --no-wandb_log when running the code):

wandb

Run the Code

In all cases go to folder src

Run GCN+JDR on Cora:

python train_model.py --dataset Cora --net GCN --data_split sparse --denoise_default GCN 

Run Rewire Baselines on Cora

DIGL

python train_model.py --dataset Cora --net GCN --data_split sparse --rewire_default ppr 

FoSR

python train_model.py --dataset Cora --net GCN --data_split sparse --rewire_default fosr 

BORF

python train_model.py --dataset Cora --net GCN --data_split sparse --rewire_default borf 

Reproduce the results of the paper:

source run_csbm_exp.sh

source run_exp_table_1.sh

source run_exp_table_2.sh

How to use this ...

...with my own GNN

Add your GNN to src/GNN_models.py and adapt the argparse in src/train_model.py accordingly. Then you can test the performance of JDR with you GNN e.g. on Cora via

python train_model.py --dataset Cora --net "your GNN" --data_split sparse --denoise_default GCN 

Consider also optimizing the hyperparameters of JDR for your specific model+datasets combination as described below.

...with my own dataset

Add your dataset to the DataLoader in src/dataset_utils.pyand adapt argparse in src/train_model.py accordingly. Since no default hyperparameters exist you need to tune them yourself. Here is a suggestion on the ranges based on our findings on the other datasets:

denoise_iterations:
  distribution: int_uniform
  max: 30
  min: 1
rewired_index_A:
  distribution: int_uniform
  max: 100
  min: 1
rewired_index_X:
  distribution: int_uniform
  max: 100
  min: 1
rewired_ratio_A:
  distribution: uniform
  max: 0.5
  min: 0
rewired_ratio_X:
  distribution: uniform
  max: 0.5
  min: 0
rewired_ratio_X_non_binary:
  distribution: uniform
  max: 1
  min: 0

Datasets

Twitch-gamers

The dataset can be downloaded from Snap.

cSBM

To create a new dataset go to folder src and run for example:

python cSBM_dataset.py --phi 0.6 --name cSBM_phi_0.6 --root ../data/ --num_nodes 5000 --num_features 2000 --avg_degree 5 --epsilon 3.25

Citation

If you find our work useful, please consider citing:

@inproceedings{linkerhagner2025joint,
  title={Joint Graph Rewiring and Feature Denoising via Spectral Resonance},
  author={Jonas Linkerh{\"a}gner and Cheng Shi and Ivan Dokmani{\'c}},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=zBbZ2vdLzH}
}