On Class Distributions Induced by Nearest Neighbor Graphs for Node Classification of Tabular Data (NeurIPS 2023)

Reference

This research software is provided as is. If you happen to use or modify this code, please remember to cite the paper:

@inproceedings{errica_class_2023,
 author = {Errica, Federico},
 title = {On Class Distributions Induced by Nearest Neighbor Graphs for Node Classification of Tabular Data},
 booktitle = {Advances in Neural Information Processing Systems},
 volume = {36},
 year = {2023}
}

How to reproduce

The steps below are used to reproduce the experiments of the paper.

Step 1) Install required packages

We assume Pytorch and Pytorch Geometric >= 2.3.0 are already installed. Then run

pip install -r requirements.txt

Step 2) Create dataset

You can prepare the datasets using the following command

pydgn-dataset --config-file DATA_CONFIGS/config_abalone.yml

(and similarly for the other datasets using the configuration files in the DATA_CONFIGS folder.)

Step 3) Launch all node classification experiments

Make sure you configure your hardware requirements in the configuration files present in the MODEL_CONFIGS folder. Then you can run

source launch_mlp_exp.sh
source launch_simpledgn_exp.sh    
source launch_gin_exp.sh    
source launch_gcn_exp.sh    

Step 4) Launch the synthetic experiments

python launch_class_separator_exp.py

This will store some checkpoints that you can load in the notebooks.

Optional Step) Launch an individual experiment (remove [--debug] to parallelize as in step 3)

pydgn-train --config-file MODEL_CONFIGS/pedalme/mlp_abalone.yml --debug

This will launch model selection and risk assessment for the MLP and compute the final scores. You can use different configuration files to launch different experiments.

Jupyter Notebooks

You can use the jupyter notebooks to inspect our qualitative experiments and animations for the theoretical results.