Bias mitigation for automated decision making systems

The increasingly widespread use of data, including personal data or relating to sensitive information, has inevitably contributed to the increase in social discrimination especially concerning automated decision systems, such as ranking systems. Previous research has made progress in identifying various types of bias that can lead to discrimination in machine learning. Techniques have been developed to reduce bias, but there are limitations in the scope of application: it is not possible to completely reset the present bias, as it is not possible to simultaneously satisfy all the principles of fairness in machine learning. We used three known datasets in the ML fairness field: the COMPAS dataset, a credit scoring dataset and a dataset with information related to drug use. Based on these data, we focused on the study of the fairness metrics and how we could improve fairness by applying certain bias mitigation techniques. Our results confirm the findings of previous research and highlight a possible process to maximize fairness, through a combined use of bias mitigation techniques.