MIIC-SR: From Complex Data to Structural Causal Models

Nadir Sella, Adam Perbost, Louis Verny

Keywords: Structural causal models, causal discovery, symbolic regression, function estimation, synthetic data generation

TL;DR: The association of a causal discovery algorithm (MIIC) and Symbolic Regression (SR) allows to estimate Structural Causal Models (SCMs) and to generate good quality synthetic data.

Abstract: Estimating the set of mathematical equations from observational data for complex systems with nonlinear relationships presents a significant challenge, particularly when the model specification is not straightforward and the causal graph is not known. We propose a comprehensive framework that estimates a Structural Causal Model (SCM) from data, requiring no prior information on the underlying causal graph. Our framework incorporates MIIC (Multivariate Information-Based Inductive Causation), a well-established causal discovery algorithm, with Symbolic Regression (SR). Our results demonstrate that the association of MIIC and Symbolic Regression shows at least comparable results in SCM estimation on the studied benchmarks with the advantage of providing an interpretable causal model.

Submission Number: 28