Explaining contributions of features towards unfairness in classifiers: A novel threshold-dependent Shapley value-based approach

A number of approaches has been proposed to investigate and mitigate unfairness in machine learning algorithms. However, as the definition and understanding of fairness may vary in different situations, the study of ethical disparities remains an open area of research. Besides the importance of analyzing ethical disparities, explainability in machine learning is also a relevant issue in Trustworthy Artificial Intelligence. Usually, both fairness and explainability analysis are based on a fixed decision threshold, which differentiates the positive cases from the negative ones according to the predicted probabilities. In this paper, we investigate how changes in this threshold can impact the fairness of predictions between protected and other groups and how features contribute towards such a measure. We propose a novel Shapley value-based approach as a tool to investigate how changes in the threshold values change the contribution of each feature towards unfairness. This gives us an ability to evaluate how fairness measures vary for different threshold values and which features have the higher (or lower) impact on creating ethical disparities. We demonstrate this using three different case studies that are carefully chosen to highlight different unfairness scenarios and features contributions. We also applied our proposal as a feature selection strategy, which contributed to decrease unfair results substantially.