Efficient Identification of Direct Causal Parents via Invariance and Minimum Error Testing

Minh Nguyen, Mert R. Sabuncu

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: causal discovery, invariance

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

Abstract: Invariant causal prediction (ICP) is a popular technique for finding direct causes (causal parents) of a target via exploiting distribution shifts. Despite its targeted search, ICP still needs to run an exponential number of tests, which significantly limits its applicability, particularly in tasks with a large number of variables to consider. Furthermore, as others have pointed out, ICP fails to identify causes when distribution shifts only affect a few variables. We propose two approaches, MMSE-ICP and fastICP, which employ an error inequality to address the identifiability problem of ICP. The inequality states that the minimum prediction error of the predictor using causal parents is the smallest among all predictors which do not use descendants. fastICP is an efficient approximation that exploits the error inequality and a heuristic to reduce the number of tests for invariance required. Our experiments on simulated and real data show MMSE-ICP and fastICP outperforming competitive baseline approaches and fastICP being much more scalable.

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: 4290