Self-Distilled Disentanglement for Counterfactual Prediction

Xinshu Li, Mingming Gong, Lina Yao

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: counterfactual prediction, disentangled representation learning, information theory, knowledge distillation.

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

TL;DR: This paper proposes a self-distilled disentanglement framework based on information theory to advance counterfactual prediction in the presence of hidden confounders.

Abstract: The advancements in disentangled representation learning significantly enhance the accuracy of counterfactual predictions by granting precise control over instrumental variables (IVs), confounders, and adjustable variables. An appealing method for achieving the independent separation of these factors is mutual information minimization (MIM), a task that presents challenges in numerous machine learning scenarios, especially within high-dimensional spaces. To circumvent this challenge, a common strategy is to re-frame the MIM problem from a problem between two high-dimensional representations to one between high-dimensional representations and low-dimensional labels based on the different dependencies of latent factors and known labels. In this paper, we first demonstrate the limitations of this approach in separating instrumental variables and confounding variables, as determined by the d-separation theory. Subsequently, we propose the Self-Distilled Disentanglement framework, referred to as $SD^2$. Grounded in information theory, it ensures theoretically sound disentangled representations without intricate mutual information estimator designs for high-dimensional representations. Our comprehensive experiments, conducted on both synthetic and real-world datasets, provide compelling evidence of the effectiveness of our approach in facilitating counterfactual inference in the presence of both observed and unobserved confounders.

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: 209