Fast Generic Interaction Detection for Model Interpretability and Compression

This repository is the official implementation of Fast Generic Interaction Detection for Model Interpretability and Compression. The code mainly relies on the MXNet package, for those who are not familiar with MXNet, we also provide a PyTorch demo for the interaction detection task.

We cast the interaction detection problem as a multi-arm bandits problem, and use the detected interactions to build a lightweight and interpretable model named ParaACE.

See detected interactions below.

Requirements

Requirements for the demo (PyTorch)

python==3.6.3

networkx==2.3

numpy==1.16.4

matplotlib==3.1.1

torch==1.1.0

If you have installed Anaconda already, then you only need to install the PyTorch additionally.

For practitioners who don't have GPU devices and for those who are not familiar with MXNet, we provide a demo for the interaction detection with UCB algorithm implemented in PyTorch.

Requirements for our main experiments (MXNet)

python==3.6.3

mxnet-cu90==1.4.1

networkx==2.3

numpy==1.16.4

matplotlib==3.1.1

pandas==0.24.2

scikit-learn==0.21.2

scipy==1.2.1

This implementation needs GPU support, please install the packages according to your own NVIDIA driver setting. For latest MXNet version, please add ctx=ctx while loading the parameters.

Make sure that you have at least 10 GB of free space available on the hard disk.

Quick Access

To have a quick access of interaction dectection, please find the demos in the demo_pytorch folder.

Datasets

Synthetic datasets

The synthetic datasets come from the functions in Tsang's paper. Run the codes in the folder gen_synthetic_data to generate the data. The data is in the csv form, the last column of which is the target.

Real datasets

The real datasets are public available.

Elevator https://www.dcc.fc.up.pt/~ltorgo/Regression/DataSets.html

Parkinsons https://archive.ics.uci.edu/ml/datasets/Parkinsons+Telemonitoring

Skillcraft http://archive.ics.uci.edu/ml/datasets/SkillCraft1+Master+Table+Dataset

Bike sharing http://archive.ics.uci.edu/ml/datasets/Bike+Sharing+Dataset

Cal housing https://www.dcc.fc.up.pt/~ltorgo/Regression/DataSets.html

Drug combination https://zenodo.org/record/4135059

Reproducibility

Different experiments are conducted in different folders.

You only need to enter the interested folder and activate your environment. You should first check all the datasets are in the regression folder.

Then, run the following command:

python main.py

It will automatically generate the result.csv file containing the comparison of different methods as well as the model files.

To visualize the detected interactions, you may open the .ipynb files and run the corresponding cells.

Comparison with LTH

Run the following command:

python LTHmain.py

You can visualize the results from the .ipynb file.

Build your own model considering interactions

Almost all machine learning models can be decomposed into interaction blocks. We additionally give a case for GP regression with additive kernel.

Please feel free to contact me if you have any questions about the reproducibility of the code. Email: tianjianzhang at link dot cuhk dot edu dot cn

Reference

If this repo is helpful in your research, please cite the following:

@inproceedings{
zhang2022fast,
title={Fast Generic Interaction Detection for Model Interpretability and Compression},
author={Tianjian Zhang and Feng Yin and Zhi-Quan Luo},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=fQTlgI2qZqE}
}