ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data (ICML 2024)

Official PyTorch implementation of "ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data (ICML 2024)"

by Xiangjian Jiang, Andrei Margeloiu, Nikola Simidjievski, Mateja Jamnik.

Installing

Local Machine

The below settings have been tested to work on: CUDA Version: 11.2 (NVIDIA-SMI 460.32.03).

<!-- Install the codebase -->
cd REPOSITORY
conda create python=3.9.0 --name protogate
conda activate protogate
pip install -r requirements.txt

<!-- Optionally, install lightgbm -->
# pip install lightgbm --install-option=--gpu --install-option="--opencl-include-dir=/usr/local/cuda/include/" --install-option="--opencl-library=/usr/local/cuda/lib64/libOpenCL.so"

Google Colab

If the environmental settings cannot work on the local machine, we also provide environmental dependecies for Google Colab. Please follow these steps:

Upload the associate notebook protogate_colab.ipynb to Google Colab
Upload the associate codebase to Google Drive
Set the project_path in notebook to the path of uploaded codebase in Google Drive
```
# Set up the path of codebase
project_path = '/path/to/codebase/'
```
Execute the notebook cells in Step1: Set up environment

Running experiments

We provide scripts for ProtoGate and benchmark methods to work on the real-world (e.g., the “meta-pam” dataset) and synthetic datasets. Below are three examples:

ProtoGate on meta-pam dataset

bash scripts/PROTOGATE/run_exp_protogate_real.sh

ProtoGate on Syn1 dataset

bash scripts/PROTOGATE/run_exp_protogate_syn.sh

The hyperparameters can be changed in the script by passing different values.

python src/run_experiment.py \
	--model 'protogate' \
	--dataset 'metabric-pam50__200' \
	--metric_model_selection total_loss \
	--lr 0.1 \
	--protogate_lam_global 0.0002 \
	--protogate_lam_local 0.001 \
	--pred_k 3 \
	--max_steps 8000 \
	--protogate_gating_hidden_layer_list 200 \
	--tags 'real-world' \
	--disable_wandb

FAQ

Q: Where to find other real-world datasets?

A: The other HDLSS datasets can be downloaded from the source website, open-source project; and the non-HDLSS datasets can be downloaded from the TabZilla benchmark.
Q: How to get the full log of training and evaluation process?

A: Please install wandb library and cancel out --disable_wandb when running the experiments.
Q: How to set up the path of files in Google Drive?

A: The only special step is to mount the Google Drive to Google Colab, and the following steps are the same as the local machine.
Q: Which license does this codebase follow?

A: This codebase will follow the Apache-2.0 license when ProtoGate is publicly available for community.

Citation

Please cite our work if this repo helps you:

@inproceedings{jiang2024protogate,
  title={ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data},
  author={Jiang, Xiangjian and Margeloiu, Andrei and Simidjievski, Nikola and Jamnik, Mateja},
  booktitle={Forty-first International Conference on Machine Learning},
  year={2024}
}