Causality is Invariance Across Heterogeneous Units
Primary Area: causal reasoning
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: DiscoModel, Layer Valuation, Counterfactual, Heterogeneous, Causal Representation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Learning a model from data for the three layers of Pearl Causal Hierarchy (PCH) (i.e., the associational, the interventional, and the counterfactual) is a central task in contemporary causal inference research, and it becomes particularly challenging for counterfactual queries. The prevailing scientific understanding is anchored in the three-step counterfactual algorithm (i.e., abduction, action, and prediction) proposed by Judea Pearl, which he considers is one of his most pivotal contributions. While this algorithm offers a theoretical solution, the absence of complete knowledge on structural causal models (SCMs) renders it highly impractical in most scenarios. To tackle the tasks of PCH, this paper introduces the DiscoModel, grounded in the core principle that "Causality is invariance across heterogeneous units." The underlying causal modeling theory of our model is \textit{Distribution-consistency Structural Causal Models} (DiscoSCMs), which extends both \textit{structural causal models} and the potential outcome framework. The former infers the selection variable on heterogeneous units, while the latter encapsulates the invariant causal relationship. DiscoModel exhibits remarkable capability for all the three layers of PCH simultaneously, providing practical and reasonable answers to important counterfactual questions (e.g., ``For a user on a certain internet platform observed with high subsidy and high retention, what if this user had not received a high subsidy in the past? Would there still be high retention now?''). To the best of our knowledge, DiscoModel is the first to provide non-trivial answers to such queries, substantiated through experiments on both simulated and real-world data.
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: 3790
Loading