CausalChat: Interactive Causal Model Development and Refinement Using Large Language Models.

Area of Science:

• Data Science
• Artificial Intelligence
• Network Science

Background:

• Causal networks are crucial for modeling complex relationships between variables across various domains.
• Existing methods for causal network construction often rely on human expertise, requiring significant domain knowledge and participation.

Purpose of the Study:

• To develop a novel approach for constructing causal networks by leveraging the knowledge embedded in large language models (LLMs).
• To introduce CausalChat, a visual analytics interface designed for interactive causal network discovery.
• To assess the efficacy of CausalChat with diverse datasets and user groups.

Main Methods:

• Utilized the causal knowledge acquired by LLMs (e.g., GPT-4) from extensive literature.
• Developed a visual analytics interface (CausalChat) enabling recursive exploration of variables.
• Translated user interactions into tailored LLM prompts for identifying causal relations, latent variables, confounders, and mediators.
• Integrated visual representations with textual explanations for enhanced understanding.

Main Results:

• Demonstrated the functionality of CausalChat across a variety of data contexts.
• User studies involving both domain experts and laypersons validated the tool's utility.
• The system successfully facilitated the construction of detailed causal networks through conversational exploration.

Conclusions:

• CausalChat offers an innovative method for causal network construction, reducing reliance on extensive human domain expertise.
• LLM-powered visual analytics presents a promising avenue for complex data relationship discovery.
• The approach is adaptable and effective for users with varying levels of domain knowledge.