METHODS FOR MINING CAUSUSITY FROM OBSERVATIONS IN ARTIFICIAL INTELLIGENCE

Abstract

The article discusses the importance of capturing causal relationships in machine learning for decision-making and evaluating real-world impact. It is noted that most current successes in machine learning are based on pattern recognition and correlation analysis; however, for more complex tasks, extracting causal relationships is necessary. The problems of explainability of predictions and causal understanding, even with the use of advanced machine learning techniques such as LIME, SHAP, TreeSHAP, DeepSHAP, and Shapley Flow, are recognized as fundamental obstacles in the development of artificial intelligence. The article briefly presents the main philosophical and mathematical concepts and definitions of causality, including counterfactuals, Bayesian networks, directed acyclic graphs, and causal formal inference. It concludes that the practical significance of data-based causal analysis consists in answering a priori formulated questions, which may reflect a hypothetical relationship between an event (a cause) and a second event (an effect), where the second event is a direct consequence of the first. A comparative analysis of the methods and main scenarios for using the Causal Discovery and Causal Inference frameworks is also carried out. Based on this analysis, it becomes possible to make assumptions about the causal structure underlying the investigated dataset and to use statistical methods to evaluate the strength and direction of such relationships. The article also discusses methods and algorithms of causal analysis and their application in real-world tasks. Representative methods are mentioned, such as constraint-based models, estimation-based models, functional causal models, (conditional) independence tests, evaluation functions, and other tools that can be used to solve the problem of extracting causal relationships from observational data. Most of these methods are implemented in open-source frameworks such as Microsoft DoWhy, Uber CausalML, causal-learn, Econ-ml, and many others, which facilitate causal analysis.