Graph Homomorphism Convolution

Proof of concept for Graph Homomorphism Convolution. http://arxiv.org/abs/2005.01214 (ICML'20)

Note: Code for left homomorphism is for our ICML'20 paper. Code for right homomorphism is our continued work.

Setup

To obtain required packages and install our modules, run

pip install -r requirements.txt

This script installs homlib and ghc.

Download and pack data

This package is not dependent on torch-geometric but we will use torch-geometric to load and pack data. Download data.zip from

https://drive.google.com/file/d/15w7UyqG_MjCqdRL2fA87m7-vanjddKNh/view?usp=sharing

and extract to data/. For example MUTAG dataset should have: data/MUTAG.graph , data/MUTAG.X, data/MUTAG.y, and data/MUTAG.folds.

Quick run with ipython

from ghc.homomorphism import hom 
from ghc.utils.data import load_data
from ghc.utils.hom import tree_list

graphs, X, y = load_data("MUTAG", "./data")
trees = tree_list(4, num_loops=1)

# right hom
print(hom(graphs[0], trees[-1]))

Run experiments

Experiment scripts are placed in the top level of this repository and named by the machine learning model. In general, a 10-fold CV score is reported. For example,

python models/mlp.py --data mutag --hom tree --size 6 
python models/mlp.py --data mutag --hom labeled_tree --size 6

The 10-fold splits for MUTAG, COLLAB, IMDBBINARY, IMDBMULTI, NCI1, PROTEINS, and PTC are taken from weihua916/powerful-gnns. The others are generated with sklearn.model_selection.StratifiedKFold at random seed 0.

Note: CPU run will cause some segmentation fault upon script exit.

Cite us as:

@InProceedings{
    nt20ghc, 
    title = {Graph Homomorphism Convolution}, 
    author = {NT, Hoang and Maehara, Takanori}, 
    booktitle = {Proceedings of the 37th International Conference on Machine Learning}, 
    pages = {7306--7316}, 
    year = {2020}, 
    editor = {Hal Daumé III and Aarti Singh}, 
    volume = {119}, 
    series = {Proceedings of Machine Learning Research}, 
    address = {Virtual}, 
    month = {13--18 Jul}, 
    publisher = {PMLR}, 
    pdf = {http://proceedings.mlr.press/v119/nguyen20c/nguyen20c.pdf},, 
    url = {http://proceedings.mlr.press/v119/nguyen20c.html}, 
}

Note: There is a bug in homlib for atlas[100:].