BLIS-Net (bi-Lipschitz Scattering Network) is a provably powerful GNN designed for graph signal classification. For further details, please refer to our AISTATS 2024 paper.
BLIS-Net consists of four modules/layers:
- The BLIS-Module
- Moment aggregation module
- Embedding/dimensionality reduction layer
- Classification layer
To accommodate a variety of workflows and tasks, we provide two equivalent implementations of the modules of BLIS-Net.
The first implementation is a pytorch implementation (code here). BLIS-Module outputs scattering features on each node, and may thus be considered as a form of message passing. This implementation may be flexibly incorporated into GNN architectures for a variety of downstream tasks. An example implentation is given in the BlisNet class from blis_legs_layer.py.
The first implementation utilizes numpy to compute the scattering moments and write them to memory (code here), after which a variety of classifiers may be trained on top of the computed scattering moments (code here).
Create a conda environment
conda create -n blis python=3.9`
conda activate blis
cd blis
pip install -e .
note: it may also be necessary to install torch-scatter
The data used in the paper is archived on Zenodo (DOI: 10.5281/zenodo.15882346) and may be downloaded from https://zenodo.org/records/15882346.
From the main project directory, download and unpack the archive into the data directory, for example:
wget -O data.zip "https://zenodo.org/records/15882346/files/data.zip?download=1"
rm -rf data
unzip data.zip
mv data_export data
rm data.zip
The archive contains three graph signal classification tasks: partly_cloudy (fMRI), synthetic, and traffic (Caltrans PeMS). See data_desc.tex on the Zenodo record for details.
A script to run the pytorch implementation is provided in scripts. From the main directory, run:
python scripts/classify_torch.py --model BlisNet --dataset synthetic --sub_dataset gaussian_pm --task_type PLUSMINUS
One script is used to first compute scattering coefficients and a second one is used to train a variety of classifiers on them. For example, one might run:
python scripts/calculate_scattering.py --scattering_type blis --wavelet_type W2 --largest_scale 4 --highest_moment 3 --dataset traffic --sub_dataset PEMS08
python scripts/classify_scattering.py --dataset=traffic --largest_scale=4 --sub_dataset=PEMS08 --scattering_type=blis --task_type=DAY --moment_list 1 --layer_list 1 2 3 --model SVC
If you have any questions or require assistance using BLIS-Net, please contact us at https://krishnaswamylab.org/contact.