Causal Machine Learning: Bridging Correlation and Causation in Artificial Intelligence

Abstract

Most machine learning systems are trained to exploit statistical correlations in data, yet many of the questions practitioners actually care about — what will happen if we intervene, or what would have happened had we acted differently — are inherently causal and cannot be answered from correlation alone. Causal machine learning integrates ideas from the statistical causal inference literature, including the potential outcomes framework and structural causal models, with modern representation learning to estimate cause-and-effect relationships from observational and experimental data. This paper surveys the foundations of causal inference, reviews algorithms for causal discovery and treatment effect estimation, and examines how deep learning has been adapted to produce causally aware representations. We discuss applications in healthcare, economics, recommendation systems, and robotics, summarize commonly used benchmarks, and present a worked example illustrating backdoor adjustment for confounding. We conclude with a discussion of open challenges, including the difficulty of verifying causal assumptions from data alone, and outline promising directions connecting causal machine learning with large-scale representation learning and out-of-distribution generalization.