aliciacurth / catenets

Sklearn-style implementations of Neural Network-based Conditional Average Treatment Effect (CATE) Estimators.

CATENets - Conditional Average Treatment Effect Estimation Using Neural Networks

Code Author: Alicia Curth ([email protected])

This repo contains Jax-based, sklearn-style implementations of Neural Network-based Conditional Average Treatment Effect (CATE) Estimators, which were used in the AISTATS21 paper 'Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms' (Curth & vd Schaar, 2021a) as well as the follow up NeurIPS21 paper "On Inductive Biases for Heterogeneous Treatment Effect Estimation" (Curth & vd Schaar, 2021b) and the NeurIPS21 Datasets & Benchmarks track paper "Really Doing Great at Estimating CATE? A Critical Look at ML Benchmarking Practices in Treatment Effect Estimation" (Curth et al, 2021).

We implement the SNet-class we introduce in Curth & vd Schaar (2021a), as well as FlexTENet and OffsetNet as discussed in Curth & vd Schaar (2021b), and re-implement a number of NN-based algorithms from existing literature (Shalit et al (2017), Shi et al (2019), Hassanpour & Greiner (2020)). We also provide Neural Network (NN)-based instantiations of a number of so-called meta-learners for CATE estimation, including two-step pseudo-outcome regression estimators (the DR-learner (Kennedy, 2020) and single-robust propensity-weighted (PW) and regression-adjusted (RA) learners), Nie & Wager (2017)'s R-learner and Kuenzel et al (2019)'s X-learner. The jax implementations in catenets.models.jax were used in all papers listed; additionally, pytorch versions of some models (catenets.models.torch) were contributed by Bogdan Cebere.

Interface

The repo contains a package catenets, which contains all general code used for modeling and evaluation, and a folder experiments, in which the code for replicating experimental results is contained. All implemented learning algorithms in catenets (SNet, FlexTENet, OffsetNet, TNet, SNet1 (TARNet), SNet2 (DragonNet), SNet3, DRNet, RANet, PWNet, RNet, XNet) come with a sklearn-style wrapper, implementing a .fit(X, y, w) and a .predict(X) method, where predict returns CATE by default. All hyperparameters are documented in detail in the respective files in catenets.models folder.

Example usage:

from catenets.models.jax import TNet, SNet
from catenets.experiment_utils.simulation_utils import simulate_treatment_setup

# simulate some data (here: unconfounded, 10 prognostic variables and 5 predictive variables)
X, y, w, p, cate = simulate_treatment_setup(n=2000, n_o=10, n_t=5, n_c=0)

# estimate CATE using TNet
t = TNet()
t.fit(X, y, w)
cate_pred_t = t.predict(X)  # without potential outcomes
cate_pred_t, po0_pred_t, po1_pred_t = t.predict(X, return_po=True)  # predict potential outcomes too

# estimate CATE using SNet
s = SNet(penalty_orthogonal=0.01)
s.fit(X, y, w)
cate_pred_s = s.predict(X)

All experiments in Curth & vd Schaar (2021a) can be replicated using this repository; the necessary code is in experiments.experiments_AISTATS21. To do so from shell, clone the repo, create a new virtual environment and run

pip install catenets # install the library from PyPI
# OR
pip install . # install the library from the local repository

# Run the experiments
python run_experiments_AISTATS.py

Options:
--experiment # defaults to 'simulation', 'ihdp' will run ihdp experiments
--setting # different simulation settings in synthetic experiments (can be 1-5)
--models # defaults to None which will train all models considered in paper,
         # can be string of model name (e.g 'TNet'), 'plug' for all plugin models,
         # 'pseudo' for all pseudo-outcome regression models

--file_name # base file name to write to, defaults to 'results'
--n_repeats # number of experiments to run for each configuration, defaults to 10 (should be set to 100 for IHDP)

Similarly, the experiments in Curth & vd Schaar (2021b) can be replicated using the code in experiments.experiments_inductivebias_NeurIPS21 (or from shell using python run_experiments_inductive_bias_NeurIPS.py) and the experiments in Curth et al (2021) can be replicated using the code in experiments.experiments_benchmarks_NeurIPS21 (the catenets experiments can also be run from shell using python run_experiments_benchmarks_NeurIPS).

The code can also be installed as a python package (catenets). From a local copy of the repo, run python setup.py install.

Note: jax is currently only supported on macOS and linux, but can be run from windows using WSL (the windows subsystem for linux).

Citing

If you use this software please cite the corresponding paper(s):

@inproceedings{curth2021nonparametric,
  title={Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms},
  author={Curth, Alicia and van der Schaar, Mihaela},
    year={2021},
  booktitle={Proceedings of the 24th International Conference on Artificial
  Intelligence and Statistics (AISTATS)},
  organization={PMLR}
}

@article{curth2021inductive,
  title={On Inductive Biases for Heterogeneous Treatment Effect Estimation},
  author={Curth, Alicia and van der Schaar, Mihaela},
  booktitle={Proceedings of the Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}


@article{curth2021really,
  title={Really Doing Great at Estimating CATE? A Critical Look at ML Benchmarking Practices in Treatment Effect Estimation},
  author={Curth, Alicia and Svensson, David and Weatherall, James and van der Schaar, Mihaela},
  booktitle={Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},
  year={2021}
}

catenets's Issues

PyTorch models are fit with propensity estimators regardless of weighting strategies used

Example (executed regardless of the value in the previous if-block at L174) :

CATENets/catenets/models/torch/slearner.py

Line 181 in bb4da46

log.info("Fit the PyTorch po_estimator with the propensity estimator")

This might appear in other places as well—maybe worth checking (@bcebere?)

PyTorch CATENets inconsistent with PyTorch APIs

The PyTorch BaseCATEEstimator, PropensityNet, RepresentationNet, BasicNet (and possibly other models based on these) have a train method that overrides the torch.nn.Module.train(), intended to set the mode to training (affecting the behaviour of Dropout, BatchNorm, etc.)

At the same time, in the actual training procedures of these models the mode is never explicitly set to training.

I would propose getting rid of the train methods in the BaseCATEEstimator interface and elsewhere, and fixing the behaviour in the corresponding fit methods (but currently don't have much time to look into the dependencies this would cause in experiments and so on).

SNet2 (jax) - UnboundLocalError: local variable 'params_best' referenced before assignment

Hi,

catenets == 0.2.1

I Get this error when trying the SNet2 or the DragonNet (jax).

TNet was working #1 with the same data and parameters.