Can AI Learn Causal Structure? Evidence from ADIA Lab's Causal Discovery Challenge

Abstract

This paper describes the ADIA Lab Causal Discovery Challenge conducted in 2024. The challenge focused on causal discovery, the task of inferring causal relationships from observational data, with an emphasis on the classification of variables according to their causal roles. Participants were tasked with identifying eight causal categories (Confounder, Collider, Mediator, Independent, Cause of X, Consequence of X, Cause of Y , and Consequence of Y) using a large synthetic database generated using known causal graphs. The competition attracted significant global participation, with 1,904 registered users submitting 3,343 solution attempts. We analyze the methodological approaches of the top-performing solutions, which achieved multiclass balanced accuracies of up to 76.70%, substantially outperforming baseline solutions (approximately 40%). An analysis of the winning methodologies showed that supervised learning approaches outperformed traditional constraint-based methods; sophisticated feature engineering and representation learning captured complex causal structures; ensemble methods combining predictions from multiple approaches yielded more robust results; and performance varied significantly across different graph structures, with hierarchical graphs being consistently easier to analyze than preferential attachment graphs. This report documents these findings and explores their implications for advancing causal discovery research.

Keywords: Causality, Causal discovery, Causal inference, Machine learning, ADIA, DAG