[AML]Leveraging LLMs for Causal Inference and Discovery
Keywords: LLM, causal discovery
Abstract: Causal inference is widely applied in the social sciences to analyze the effects of a specific treatment. Causal inference tools rely on uncovering the underlying causal graph in advance, a process known as causal discovery. Traditionally, constructing causal graphs has depended on expert domain knowledge; however, the rich knowledge embedded in large language models (LLMs) offers a promising alternative. Nevertheless, LLMs alone perform poorly in inferring complete causal graphs, primarily because they fail to account for the directed acyclic nature of causal graphs. To address this limitation, we propose a novel approach that combines LLMs with statistical causal discovery algorithms to better leverage the expert-like capabilities of LLMs. Experimental results demonstrate that the proposed method significantly improves the accuracy of causal ordering and effectively reduces errors in downstream causal effect estimation tasks.
Submission Number: 24
Loading