Causal Structure Learning Supervised by Large Language Model

Taiyu Ban, Lyuzhou Chen, Derui Lyu, Xiangyu Wang, Huanhuan Chen

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, Large Language Model, Causal structure learning, Bayesian Network structure learning

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 paradigm of using Large Language Model to supervise causal discovery and reaches sota performance surpassing data-based CSL.

Abstract:

Causal discovery from observational data is pivotal for deciphering complex relationships. While Causal Structure Learning (CSL) aims to extract causal Directed Acyclic Graphs (DAGs), its efficacy is hampered by the expansive DAG space and data sparsity. The advent of Large Language Models (LLMs) presents a novel avenue, given their aptitude in causal reasoning, thereby constraining CSL with knowledge-based causal inference.
A pioneering study integrated LLMs into CSL, achieving notable results in several real-world DAGs. Yet, it faced pitfalls such as erroneous LLM inferences and the inefficacy of ancestral constraints.
In response, we introduce the Iterative LLM Supervised CSL (ILS-CSL) framework. This approach seamlessly merges LLM-based causal inference with CSL, iteratively refining the causal DAG based on LLM feedback.
Given LLM's shortness in distinguishing indirect causality form the direct, ILS-CSL is still capable to offer constraints on direct causality that are more powerful than the indirect, by integrating statistical dependencies indicated by data.
Moreover, the prior errors are significantly reduced while using identical LLM resources.
Our evaluations on eight real-world datasets confirm ILS-CSL's dominance, establishing a new benchmark in CSL performance.

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: 1732