Counterfactual Regression using Balancing Neural Networks as developed by Johansson, Shalit & Sontag (2016) and Shalit, Johansson & Sontag (2016). cfrnet is implemented in Python using TensorFlow and NumPy.
The core components of cfrnet, i.e. the TensorFlow graph, is contained in cfr_net.py. A simple training script is contained in cfr_train_simple.py. This file takes many flag parameters, all of which are listed at the top of the file.
In the root directory there is an example script called run_simple.sh which calls the python script cfr_train_simple.py. This example trains a counterfactual regression model on a single realization of the simulated IHDP data (see references), contained in data/ihdp_sample.csv. It creates a folder called results/single_<config & timestamp>/ which contains 5 files:
- config.txt - The configuration used for the run
- log.txt - A log file
- loss.csv - The objective, factual, counterfactual and imbalance losses over time
- y_pred.csv - The predicted factual and counterfactual outputs for all units)
- results.npz - A numpy array file with the fields "pred" and "loss" which contains the same output as the previous two files.
The data (.csv) file has the following columns: treatment (0 or 1), y_factual, y_cfactual, mu0, mu1, x_1, …, x_d. mu0 and mu1 are not used in training (they are the true simulated outcomes under control and treatment, without noise).
Uri Shalit, Fredrik D. Johansson & David Sontag. Bounding and Minimizing Counterfactual Error, arXiv:1606.03976 Preprint, 2016
Fredrik D. Johansson, Uri Shalit & David Sontag. Learning Representations for Counterfactual Inference. 33rd International Conference on Machine Learning (ICML), June 2016.