Combining causal analysis and machine learning to predict the effects of interventions

This paper examines the integration of causal analysis (causality) and machine learning methods to accurately predict the effects of interventions. The first part introduces the rationale for the importance of the causal approach when classical statistical models and purely associative ML methods face problems of hidden factors and incorrect extrapolation of results. The second part discusses the basic theoretical concepts of causal graphs, do-operator, intervening and counterfactual distributions, and the role of identifiability assumptions in the presence of unobserved confounders. Next, methods for integrating causality and machine learning - causal supervised learning (to deal with spurious correlations and increase robustness to distributional shifts), causal generative modeling (with a focus on generating counterfactual data), and other state-of-the-art approaches (causal model explanation, causal fairness, causal reinforcement learning) - are discussed in detail. It is shown how such methods can better account for the real-world structure of the data and produce more reliable predictions, especially in heterogeneous environments. The results can be applied to medicine, economics, social sciences, and other fields where it is important to accurately predict the effects of potential interventions.