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

TL;DR

This work adopts a different approach: leveraging the causal knowledge that large language models, such as OpenAI's GPT-4, have learned by ingesting massive amounts of literature to construct causal networks through conversation.

Abstract

Causal networks are widely used in many fields to model the complex relationships between variables. A recent approach has sought to construct causal networks by leveraging the wisdom of crowds through the collective participation of humans. While this can yield detailed causal networks that model the underlying phenomena quite well, it requires a large number of individuals with domain understanding. We adopt a different approach: leveraging the causal knowledge that large language models, such as OpenAI's GPT-4, have learned by ingesting massive amounts of literature. Within a dedicated visual analytics interface, called CausalChat, users explore single variables or variable pairs recursively to identify causal relations, latent variables, confounders, and mediators, constructing detailed causal networks through conversation. Each probing interaction is translated into a tailored GPT-4 prompt and the response is conveyed through visual representations which are linked to the generated text for explanations. We demonstrate the functionality of CausalChat across diverse data contexts and conduct user studies involving both domain experts and laypersons.

Figures (12)

• Figure 1: Workflow of our ChatGPT-powered Causal Auditor. (1) (Optional) algorithmic discovery of the initial (raw) causal model. (2) Query-driven ChatGPT-based edge commentary. (3) Analyst-initiated model refinement informed by the outcomes of steps 1 and 2. (4) Data upload for newly introduced variables and relations (if available). (SLA stands for Structure Learning Algorithm and SEM stands for Structural Equation Modeling).
• Figure 2: Causal Debate Chart for the relation Percent Fair or Poor Health - Life Expectancy, presenting an overwhelming belief that the former is the cause of the latter.
• Figure 3: Causal Relation Environment Chart for the relation Percent Fair or Poor Health - Life Expectancy. The intensity of red and green encodes the strength of the mediators and covariates (weak, medium, strong), and the color of the cause and effect variables have the same interpretation as those in Fig. \ref{['fig:cs-dbt']}; in this specific case they are grey.
• Figure 4: Causal Relation Environment Chart for two level combinations of the relation Percent Fair or Poor Health - Life Expectancy. The up and down arrows show the appropriate signs of the mediators.
• Figure 5: Causal Relation Environment Chart for an improbable level combination of the relation Percent Fair or Poor Health - Life Expectancy, namely one where both variables have positive levels. The up arrows in the mediators show how this improbable combination might be achieved, in form of interventions on the mediators in the direction of the arrows.