Feature selection is a critical part of creating machine learning models, particularly for touchy areas such as HR, finance, and healthcare. The literature shows that many current methods only review the results based on accuracy while ignoring both fairness and interpretability. This study presents a feature selection method which fuses Mutual Information for assessing feature relevance, while imposing fairness constraints by measuring the Equal Opportunity Difference (EOD) metric. The aim, of this study, is to maximize accuracy, fairness, and interpretability. The methodology was tested on the IBM HR Analytics dataset for six models (Fair Logistic Regression, Explainable Boosting Machines (EBMs), XGBoost, Random Forest, SVM and KNN). Fair Logistic Regression, provided the best overall results with 96% accuracy and little fairness bias (EOD = 0.005). XGBoost and Random Forest had good predictive reliability, evidence of disparity in fairness measures. Moreover, Explainable Boosting Machines (EBMs) met accuracy standards, and have good interpretability, so these models could be used in intervention areas where transparency of the model/construction process is required.
omar shakir (Mon,) studied this question.