Since Faithfulness Fails: The Performance Limits of Neural Causal Discovery

Mateusz Olko, Mateusz Gajewski, Joanna Wojciechowska, Łukasz Kuciński, Mikołaj Morzy, Piotr Sankowski, Piotr Miłoś

Keywords: causal discovery, faithfulness assumption, neural networks

TL;DR: We show that the faithfulness assumption limits the structure discovery accuracy of neural causal discovery method on standard benchmarks.

Abstract:

Neural causal discovery methods have recently improved in terms of scalability and computational efficiency.
However, there are still opportunities for improving their accuracy in uncovering causal structures.
We argue that the key obstacle in unlocking this potential is the faithfulness assumption, commonly used by contemporary neural approaches. We show that this assumption, which is often not satisfied in real-world or synthetic datasets, limits the effectiveness of existing methods. We evaluate
the impact of
faithfulness violations both qualitatively and quantitatively and provide a unified evaluation framework to facilitate further research.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 10102