Causal CSSE: Integrating Counterfactuals and Causality in the Explanation of Machine Learning Models

Abstract

Machine learning methods are widely used in various industries and research sectors, including health, agriculture, and law, to support decision-making. However, many of these models are inherently uninterpretable, necessitating the development of explainability methods. Among those methods, counterfactual methods aim to explain the decisions of black-box models by making minimal modifications to an instance’s attributes, thereby identifying the necessary changes to alter its classification. This article examines the relationship between counterfactuals and causality, emphasizing the significance of understanding causality for achieving more consistent explainability. We present the Causal CSSE method, which incorporates aspects of causality into the counterfactual CSSE (Social Explanations for Classification Models) method, providing more realistic and applicable explanations. Causal CSSE uses an ad-hoc genetic model to generate counterfactual examples, considering the causality between attributes to improve the interpretability and applicability of the generated explanations. This enhancement aims to provide more realistic and valuable counterfactual answers, thereby contributing to more effective solutions for real-world problems. This project is available at https://github.com/virion1996/causal_csse .