Jiongli Zhu, Babak Salimi.
The pervasive integration of machine learning (ML) across various sectors has underscored the critical challenge of addressing inherent biases in ML models. These biases not only undermine the models’ fairness and accuracy but also have significant real-world consequences. Traditional approaches to mitigating these biases often fail to address their root causes, leading to solutions that may superficially seem fair but do not tackle the underlying problems. This review paper explores the role of causal modeling in enhancing data cleaning, preparation, and quality management for ML. By analyzing existing research, we demonstrate how causal reasoning can effectively identify and rectify data biases, thus improving the fairness and accuracy of ML models. We advocate for the increased adoption of causal approaches in these processes, emphasizing their potential to significantly enhance the integrity and reliability of data-driven technologies.
Paper
