Primary Submission Category: Causal Discovery
LLM-Augmented Human-in-the-Loop Causal Discovery
Authors: Chi Zhang, Scott Mueller, Rumen Iliev, Laura Libby, Laith Ulaby, Candice Hogan,
Presenting Author: Chi Zhang*
Causal structure discovery is a challenging problem. Some core challenges include the limitation of data and scalability to large models. Data alone are almost never sufficient for learning one single causal structure, and real-world data can be noisy. Expert knowledge is helpful in filling in the blanks and correcting the errors, but it becomes cognitively difficult for human experts when the graphs are large. In this work, we address those challenges by developing a novel causal discovery approach that combines constraints from the data, knowledge from large-language models (LLMs), and human expertise. This approach aims to improve the accuracy of the discovered causal structure by iteratively applying information from the three sources. It lowers the burden of human experts by using LLMs to infer causal relationships at the same time. We derive theorems on the accuracy of the proposed algorithm, and empirically evaluate the performance through comparisons with baseline algorithms on synthetic datasets.