Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods
Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods
Derek Lim*, Felix Hohne*, Xiuyu Li*, Sijia Linda Huang, Vaishnavi Gupta, Omkar Bhalerao, Ser-Nam Lim
Published at NeurIPS 2021
Here are codes to load our proposed datasets, compute our measure of homophily, and train various graph machine learning models in our experimental setup. We include an implementation of the new graph neural network LINKX that we develop.
Organization
main.py contains the main full batch experimental scripts.
main_scalable.py contains the minibatching experimental scripts.
parse.py contains flags for running models with specific settings and hyperparameters.
dataset.py loads our datasets.
models.py contains implementations for graph machine learning models, though C&S (correct_smooth.py, cs_tune_hparams.py) are in separate files. Running several of the GNN models on larger datasets may require at least 24GB of VRAM. Our LINKX model is implemented in this file.
homophily.py contains functions for computing homophily measures, including the one that we introduce in our_measure.
experiments/ contains the bash files to reproduce full batch experiments.
scalable_experiments/ contains the bash files to reproduce minibatching experiments.
wiki_scraping/ contains the Python scripts to reproduce the "wiki" dataset by querying the Wikipedia API and cleaning up the data.
Datasets
As discussed in the paper, our proposed datasets are "genius", "twitch-gamer", "fb100", "pokec", "wiki", "arxiv-year", and "snap-patents", which can be loaded by load_nc_dataset in dataset.py by passing in their respective string name. Many of these datasets are included in the data/ directory, but wiki, twitch-gamer, snap-patents, and pokec are automatically downloaded from a Google drive link when loaded from dataset.py. The arxiv-year dataset is downloaded using OGB downloaders. load_nc_dataset returns an NCDataset, the documentation for which is also provided in dataset.py. It is functionally equivalent to OGB's Library-Agnostic Loader for Node Property Prediction, except for the fact that it returns torch tensors. See the OGB website for more specific documentation. Just like the OGB function, dataset.get_idx_split() returns fixed dataset split for training, validation, and testing.
When there are multiple graphs (as in the case of fb100), different ones can be loaded by passing in the sub_dataname argument to load_nc_dataset in dataset.py. In particular, fb100 consists of 100 graphs. We only include ["Amherst41", "Cornell5", "Johns Hopkins55", "Penn94", "Reed98"] in this repo, although others may be downloaded from the internet archive. In the paper we test on Penn94.
References
The datasets come from a variety of sources, as listed here:
- Penn94. Traud et al 2012. Social Structure of Facebook Networks
- pokec. Leskovec et al. Stanford Network Analysis Project
- arXiv-year. Hu et al 2020. Open Graph Benchmark
- snap-patents. Leskovec et al. Stanford Network Analysis Project
- genius. Lim and Benson 2020. Expertise and Dynamics within Crowdsourced Musical Knowledge Curation: A Case Study of the Genius Platform
- twitch-gamers. Rozemberczki and Sarkar 2021. Twitch Gamers: a Dataset for Evaluating Proximity Preserving and Structural Role-based Node Embeddings
- wiki. Collected by the authors of this work in 2021.
Installation instructions
- Create and activate a new conda environment using python=3.8 (i.e.
conda create --name non-hom python=3.8) - Activate your conda environment
- Check CUDA version using
nvidia-smi - run
bash install.sh cu110, replacing cu110 with your CUDA version (CUDA 11 -> cu110, CUDA 10.2 -> cu102, CUDA 10.1 -> cu101). We tested on Ubuntu 18.04, CUDA 11.0.
Running experiments
- Make sure a results folder exists in the root directory.
- Our experiments are in the
experiments/andscalable_experiments/directory. There are bash scripts for running methods on single and multiple datasets. Please note that the experiments must be run from the root directory, e.g. (bash experiments/mixhop_exp.sh snap-patents). For instance, to run the MixHop experiments on arxiv-year, use:
bash experiments/mixhop_exp.sh arxiv-year
To run LINKX on pokec, use:
bash experiments/linkx_exp.sh pokec
To run LINK on Penn94, use:
bash experiments/link_exp.sh fb100 Penn94
To run GCN-cluster on twitch-gamers, use:
bash scalable_experiments/gcn_cluster.sh twitch-gamer
To run LINKX minibatched on wiki, use
bash scalable_experiments/linkx_exp.sh wiki
To run LINKX on Geom-GCN with full hyperparameter grid on chameleon, use
bash experiments/linkx_tuning.sh chameleon
It seems that there are 3 files to be downloaded. While in in github page:
There are only 2 files with different names.
126 Dec 3, 2022
134 Dec 16, 2022
2 Dec 20, 2021
37 Dec 8, 2022
51 Nov 18, 2022
3 Jun 22, 2022
130 Dec 2, 2022
26 Nov 4, 2022
93 Nov 6, 2022
45 Dec 12, 2022
132 Dec 18, 2022
184 Dec 11, 2022
25.1k Jan 2, 2023
5 Sep 16, 2022
190 Jan 3, 2023
107 Jan 4, 2023
44 Sep 15, 2022
373 Jan 2, 2023
128 Dec 8, 2022