Review of Causal Inference Deep Learning Methods for ITE Estimation with Automatic Hyperparameter Optimization

Tensorflow implementation of methods presented in:

Andrei Sirazitdinov, Marcus Buchwald, Jürgen Hesser, and Vincent Heuveline "Review of Deep Learning Methods for Individual Treatment Effect Estimation with Automatic Hyperparameter Optimization", 2022. Submitted to IEEE Transactions on Neural Networks and Learning.

Contact: andrei.sirazitdinov@medma.uni-heidelberg.de, marcus.buchwald@medma.uni-heidelberg.de

Currently the following causal inference methods present in the library: TLearner, SLearner, RLearner, XLearner, TARNet, CFR-Wasserstein, CFR-Weighted, CFR-MMD, DragonNet, TEDVAE, CEVAE, GANITE, DKLITE.

Requirements:

1. Python 3.9
2. Tensorflow 2.11
3. Tensorflow Probability 0.19.0
4. keras-tuner 1.1.3
5. protobuf<=3.20.x
6. pandas
7. scikit-learn
8. Numpy

To run the code use:

python main.py [-h] [--model-name MODEL_NAME] [--ipm-type IPM_TYPE] [--dataset-name DATASET_NAME] [--num NUM]

Example:
python main.py --model-name TARnet --dataset-name ihdp_b --num 100 --ipm-type None

Alternatively use Jupyter notebook.

See the full list of available models and datasets in main.py.

The file hyperparameters.py contains hyperparameters such as batch size or learning rate for the presented models.

Note, that our code performs hyperparameter search at first execution for each method to find the other hyperparameters.

We output PEHE or policy risk for each sub dataset and after training on all datasets an average PEHE or policy risk with 95% confidence interval.