Neural Causal Structure Discovery from Interventions

Nan Rosemary Ke, Olexa Bilaniuk, Anirudh Goyal, Stefan Bauer, Hugo Larochelle, Bernhard Schölkopf, Michael Curtis Mozer, Christopher Pal, Yoshua Bengio

Authors that are also TMLR Expert Reviewers: ~Yoshua_Bengio1

Abstract: Recent promising results have generated a surge of interest in continuous optimization methods for causal discovery from observational data. However, there are theoretical limitations on the identifiability of underlying structures obtained solely from observational data. Interventional data, on the other hand, provides richer information about the underlying data-generating process. Nevertheless, extending and applying methods designed for observational data to include interventions is a challenging problem. To address this issue, we propose a general framework based on neural networks to develop models that incorporate both observational and interventional data. Notably, our method can handle the challenging and realistic scenario where the identity of the intervened upon variable is unknown. We evaluate our proposed approach in the context of graph recovery, both de novo and from a partially-known edge set. Our method achieves strong benchmark results on various structure learning tasks, including structure recovery of synthetic graphs as well as standard graphs from the Bayesian Network Repository.

Certifications: Expert Certification

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

We express our sincere gratitude to the reviewers for their thorough and valuable feedback, which has significantly contributed to the improvement of our paper. In this updated version, we have implemented the following edits:

1. Data Generation Process: We have provided a more comprehensive description of the data generation process, particularly regarding interventions. This information can now be found in Section 5 (Synthetic Data).

2. Baseline Method Comparisons: We have included discussions on the selection of competing methods for the experiments and how the comparisons between different methods were conducted. Specifically, details about the data provided to the various baseline methods can be found in the paragraph titled "Baseline Comparisons" in Section 6.

3. Typos and Missing References: We have rectified the identified typos and included the missing references as pointed out by the reviewers.

4. Section 6.5 (previously Section 5.5) has been updated to provide a clearer explanation of the concept of known and unknown interventional generation.

5. Runtimes Update: We have revised the runtimes of our methods and the baseline methods, and this information can be found in Appendix Section A.11.

6. Section 6.4 now provides an explanation for the omission of competitor methods from Table 1 in Table 4.

7. Loss Function in Section 4.1: We have improved the mathematical clarity of Section 4.1 by including a detailed explanation of the loss function.

Once again, we extend our appreciation to the reviewers for their invaluable feedback, which has significantly contributed to enhancing the quality of our paper.

Assigned Action Editor: ~Daniel_M_Roy1

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Submission Number: 992