Network structure impact on graph neural networks

Papers

This code is the basis for our

• The Interplay between Communities and Homophily in Semi-supervised Classification Using Graph Neural Networks
• On the Impact of Communities on Semi-supervised Classification Using Graph Neural Networks

Citation

@article{Hussain2021,
author = {Hussain, Hussain and Duricic, Tomislav and Lex, Elisabeth and Helic, Denis and Kern, Roman},
doi = {10.1007/s41109-021-00423-1},
file = {:home/hhussain/literature/s41109-021-00423-1.pdf:pdf},
issn = {2364-8228},
journal = {Applied Network Science},
keywords = {Community structure,Graph neural networks,Homophily,Semi-supervised learning,community structure,graph neural networks,homophily,semi-},
pages = {1--26},
publisher = {Springer International Publishing},
title = {{The interplay between communities and homophily in semi ‑ supervised classification using graph neural networks}},
url = {https://doi.org/10.1007/s41109-021-00423-1},
year = {2021}
}
@inproceedings{hussain2020impact,
  title={On the Impact of Communities on Semi-supervised Classification Using Graph Neural Networks},
  author={Hussain, Hussain and Duricic, Tomislav and Lex, Elisabeth and Kern, Roman and Helic, Denis},
  booktitle={International Conference on Complex Networks and Their Applications},
  pages={15--26},
  year={2020},
  organization={Springer}
}

Dependencies

python >= 3.7

Please install PyTorch and PyG, then run pip install -r requirements.txt.

Datasets

First, run python -m src.data.dataset_handle to download and transform the datasets. This works for Cora, Citeseer, WebKB and Pubmed without hassle. Needs some tweaks to work on other datasets (to be fixed/explained).

Generating synthetic graphs from the original ones

Configuration model

Eliminates community structure while keeping the degree sequence.

Use python -m src.data.create_configuration_model

Stochastic block model

Eliminates the skew in the degree distribution (approaches a binomial distribution) while aiming to preserve the community structure using Louvain method for community detection.

Use python -m src.data.create_configuration_model

Erdős–Rényi model

Eliminates the community structure and turns the degree distribution into a binomial distribution. The only preserved properties are the node sequence (number, identity and features) and an approximate edge density.

Use python -m src.data.create_random_graph

Hyperparameter optimization

Run python -m src.hyperparam_search with the suitable parameters. You can modify the ranges within the python file. Stores validation results in reports/results/eval/ which will be necessary to run train.py.

Training and evaluating

Run python -m src.train with the suitable parameters. This file uses the resutls from the previous step and stores the evaluation resutls in reports/results/test_acc.

Computing the uncertainty coefficient

View the notebooks/Uncertainty coefficient.ipynb for details