This repository contains the implementation code for paper:
Auto IV: Counterfactual Prediction via Automatic Instrumental Variable Decomposition
Junkun Yuan, Anpeng Wu, Kun Kuang, Bo Li, Runze Wu, Fei Wu, Lanfen Lin
ACM Transactions on Knowledge Discovery from Data (TKDD), 2022
[arXiv]
Instrumental Variable (IV) is a powerful tool for causal inference, but it is hard to find/pre-define valid IVs. We propose an Automatic Instrumental Variable decomposition (AutoIV) algorithm to generate IV representations from observed variables through mutual information constraints for IV-based counterfactual prediction.
You may need to build suitable Python environment for the experiments.
The following package versions are recommened.
- python 3.6
- tensorflow-gpu 1.15.0
Device:
- GPU with VRAM > 3GB (strictly).
- Memory > 4GB.
- Configure run.sh file.
- Run the code with command:
nohup sh run.sh > run.txt 2>&1 &
- Your may check the results in the following path:
| Information | Path to check | Note |
|---|---|---|
| Generated synthetic data | data/dataset/dataset-train_data_num/ |
x: treatment; y: structural response; ye: true outcome. |
| Generated IV representation | data/dataset/autoiv-dataset/autoiv-dataset-train_data_num-rep_rep_dim/data/ |
col 1 (x): treatment; col 2 (x_pre): treatment predicted by treatment network; col 3 (y): true outcome; col 4 (y_pre): outcome predicted by outcome network; col 5~{4+ rep_dim}: IV representation Z;col {5+ rep_dim}~{4+2*rep_dim}: confounder representation C. |
| Training details of AutoIV | AutoIV-results/dataset-train_data_num-date/ |
Trace of Loss and MSE error of training, validation, and test data during training. |
- [01/13/2022] We have updated our arXiv version. See https://arxiv.org/abs/2107.05884.
- [01/09/2022] Our work is published on TKDD 2022. See https://dl.acm.org/doi/10.1145/3494568.
If you find our code or idea useful for your research, please consider citing our work.
@article{yuan2022auto,
title={Auto IV: Counterfactual Prediction via Automatic Instrumental Variable Decomposition},
author={Yuan, Junkun and Wu, Anpeng and Kuang, Kun and Li, Bo and Wu, Runze and Wu, Fei and Lin, Lanfen},
journal={ACM Transactions on Knowledge Discovery from Data (TKDD)},
volume={16},
number={4},
pages={1--20},
year={2022},
publisher={ACM New York, NY}
}If you have any questions, feel free to contact us through email (yuanjk@zju.edu.cn or anpwu@zju.edu.cn) or GitHub issues. Thanks!