Adversarial Learning of Decomposed Representations for Treatment Effect Estimation

Xiangyu Zheng, Guogang Tian, Sen Wang, Zhixiang Huang

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Causal Inference, Decomposed Representations, Individual Treatment Effect, Observational Data, Representation Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose the ADR algorithm to learn the decomposed representations for the ITE estimation, which can be applied to both categorical and numerical treatments and the disentanglement is assured by both theoretical analyses and empirical results.

Abstract:

Estimating the Individual-level Treatment Effect (ITE) from observational data is an important issue both theoretically and practically. Including all the pre-treatment covariates for prediction is unnecessary and may aggravate the issue of data unbalance.
While the confounders (C) are necessary, there are some covariates that only affect the treatment (instrumental variables, I), and some only affect the outcome (adjustment variables, A). Theoretical analyses show that including extra information in I may increase the variance lower bound and hence should be discarded. To facilitate the decomposed representation learning for the ITE estimation, we provide a rigorous definition of {I, C, A} in terms of the causal graph and prove that such decomposition is identifiable from observational data. Under the guidance of such theoretical justification, we propose an effective ADR algorithm to learn the decomposed representations and simultaneously estimate the treatment effect by introducing adversarial modules to constrain the independence and conditional independence relations. Our proposed algorithm can be applied to both categorical and numerical treatments and the disentanglement is assured by both theoretical analyses and empirical results. Experimental results on both synthetic and real data show that the ADR Algorithm is advantageous compared to the state-of-the-art methods. Theoretical analyses also provide a path to further explore the issue of decomposed representation learning for ITE estimation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2464