A preprocessing Shapley value-based approach to detect relevant and disparity prone features in machine learning

Decision support systems became ubiquitous in every aspect of human lives. Their reliance on increasingly complex and opaque machine learning models raises transparency and fairness concerns with respect to unprivileged groups of people. This motivated several efforts to estimate importance of features towards the models' performance and to detect unfair/disparate decisions. The latter is often dealt with by means of fairness metrics that rely on performance metrics with respect to predefined features that are considered protected (salient features such as age, gender, ethnicity, etc.) and/or sensitive (such as education, /occupation, banking information). However, such an approach is subjective (as fairness metrics depend on the choice features), there may be other features that lead to unfair (disparate) decisions and that may ask for suitable interpretations.