Detecting Simpson’s Paradox: A Step Towards Fairness in Machine Learning

In the last two decades, artificial intelligence (AI) and machine learning (ML) have grown tremendously. However, understanding and assessing the impacts of causality and statistical paradoxes are still some of the critical challenges in their domains. Currently, these terms are widely discussed within the context of explainable AI (XAI) and algorithmic fairness. However, they are still not in the mainstream AI and ML application development scenarios. In this paper, first, we discuss the impact of Simpson’s paradox on linear trends, i.e., on continuous values, and then we demonstrate its effects via three benchmark training datasets used in ML. Next, we provide an algorithm for detecting Simpson’s paradox. The algorithm has experimented with the three datasets and appears beneficial in detecting the cases of Simpson’s paradox in continuous values. In future, the algorithm can be utilized in designing a certain next-generation platform for fairness in ML.