Repository for NeurIPS 2020 paper: Subgraph Neural Networks
Authors: Emily Alsentzer*, Sam Finlayson*, Michelle Li, Marinka Zitnik
To use SubGNN, do the following:
- Install the environment
- Prepare data
- Modify
PROJECT_ROOT
inconfig.py
- Modify the appropriate
config.json
file - Train and evaluate SubGNN
We provide a yml file containing the necessary packages for SubGNN. Once you have conda installed, you can create an environment as follows:
conda env create --file SubGNN.yml
Prepare data for SubGNN by either (1) downloading our provided datasets or following the steps in the prepare_dataset
folder README to (2) generate synthetic datasets or (3) format your own data.
Real-World Datasets: We are releasing four new real-world datasets: HPO-NEURO, HPO-METAB, PPI-BP, and EM-USER. You can download these files from Dropbox here. You should unzip the folder and set the PROJECT_ROOT
in config.py
to the path where you downloaded the data (e.g. /PATH/TO/SubGNN_data
).
Synthetic Datasets: We also provide a script to generate the DENSITY, CORENESS, COMPONENT, and CUTRATIO synthetic graphs featured in our paper. See the README in the prepare_dataset
folder for more information on how to generate these synthetic datasets.
Your Own Data: To use your own data with SubGNN, you will need an edge list file containing the edges of the base graph and a file containing the node ids of the subgraphs, their labels, and whether they are in the train/val/test splits. Then you will need to generate node embeddings and precompute similarity metrics. For more info on how to do this, refer to the README in the prepare_dataset
folder.
To train SubGNN, you should first specify your project directory via PROJECT_ROOT
in config.py
if you haven't already. This directory should include folders containing all datasets and will ultimately contain all tensorboard folders with model outputs. Then, modify the config.json
file for the appropriate dataset to set the tensorboard output directory and the hyperparameter search ranges, including which SubGNN channels (neighborhood, structure, or position) to turn on. To learn more about the hyperparameters, go to the README
in the config_files
folder. Finally, train the model via the following:
cd SubGNN
python train_config.py -config_path config_files/hpo_metab/metab_config.json
The model and asssociated hyperparameters will be saved in the tensorboard directory specified by tb_dir
and tb_name
in the config file. We use the hpo_metab
dataset as as example, but you can easily run any of the datasets by passing in the appropriate config file. Note that, while you can also train the model via train.py
, we highly recommend using train_config.py
instead.
Once you have trained SubGNN and selected the best hyperparameters on the validaation set, run the test.py
script to re-train the model on 10 random seeds and evaluate on the test set:
cd SubGNN
python test.py \
-task hpo_metab \
-tb_dir NAME_OF_TENSORBOARD_FOLDER \
-tb_name NAME_OF_RUN_TYPE
-restoreModelPath PATH/TO/MODEL/LOCATION/WITH/BEST/HYPERPARAMETERS
Note that the restoreModelPath
directory should contain a .ckpt
file and a hyperparams.json
file. This command will create a tensorboard directory at PROJECT_ROOT/tb_dir/tb_name
where tb_dir
and tb_name
are specified by the input parameters. The test performance on each random seed will be saved in test_results.json
files in folders in this tensorboard directory. The experiment_results.json
file summarizes test performance across all random seeds.
You can also evaluate the model on a single random seed. You can use train.py
with the -noTrain
and -runTest
flags to restore a specific model and evaluate on test data. The results will be printed to the console.
cd SubGNN
python train.py \
-task hpo_metab \
-noTrain \
-runTest \
-no_save \
-restoreModelPath PATH/TO/SAVED/MODEL \
-restoreModelName CHECKPOINT_FILE_NAME.ckpt
@article{alsentzer2020subgraph,
title={Subgraph Neural Networks},
author={Alsentzer, Emily and Finlayson, Samuel G and Li, Michelle M and Zitnik, Marinka},
journal={Proceedings of Neural Information Processing Systems, NeurIPS},
year={2020}
}
Please open an issue or contact [email protected] with any questions.