Not Logged In

CounterFactual Regression with Importance Sampling Weights

Full Text: icml2018-counterfactual-regression.pdf PDF
Other Attachments: CausalML2018_poster_final.pdf [Poster] PDF

Perhaps the most pressing concern of a patient recently diagnosed with cancer is her life expectancy for various treatment options. Answering such questions requires estimating the unobserved (i.e., counterfactual) outcomes of the treatments that were not administered for each patient in the training data. This “counterfactual challenge” applies not only to healthcare, but also to other fields such as education, economics, etc. This paper extends the work of Shalit et al. (2017) for estimating Individualized Treatment Effect (ITE) in two directions: modifying (i) the objective function by adding importance sampling weights, and (ii) the procedure for finding the best set of model hyperparameters. Our evaluation on the synthetic datasets from the 2018 Atlantic Causal Inference Conference Data Challenge demonstrated significantly better performance of the proposed weighting scheme over that of (Shalit et al., 2017).

Citation

N. Hassanpour, R. Greiner. "CounterFactual Regression with Importance Sampling Weights". ICML 2018 CausalML: Workshop on Machine Learning for Causal Inference, Counterfactual Prediction, July 2018.

Keywords:  
Category: In Workshop

BibTeX

@misc{Hassanpour+Greiner:CausalML18,
  author = {Negar Hassanpour and Russ Greiner},
  title = {CounterFactual Regression with Importance Sampling Weights},
  booktitle = { ICML 2018 CausalML: Workshop on Machine Learning for Causal
    Inference, Counterfactual Prediction},
  year = 2018,
}

Last Updated: July 10, 2018
Submitted by Negar Hassanpour

University of Alberta Logo AICML Logo