Bias in Algorithms: On the trade-off between accuracy and fairness

Within the discussion on bias in algorithmic selection, fairness interventions are increasingly becoming a popular means to generate more socially responsible outcomes. The paper uses a modified framework based on Rambachan et. al. (2020) to empirically investigate the extent to which bias mitigation techniques can provide a more socially responsible outcome and prevent bias in algorithms. In using the algorithmic auditing tool AI Fairness 360 on a synthetically biased dataset, the paper applies different bias mitigation techniques at the preprocessing, inprocessing and postprocessing stage of algorithmic selection to account for fairness. The data analysis has been aimed at detecting violations of group fairness definitions in trained classifiers. In contrast to previous research, the empirical analysis focusses on the outcomes produced by decisions and the incentives problems behind fairness. The paper showed that binary classifiers trained on synthetically generated biased data while treating algorithms with bias mitigation techniques leads to a decrease in both social welfare and predictive accuracy in 43% of the cases tested. The results of our empirical study demonstrated that fairness interventions, which are designed to correct for bias often lead to worse societal outcomes. Based on these results, we propose that algorithmic selection involves a trade-between accuracy of prediction and fairness of outcomes. Furthermore, we suggest that bias mitigation techniques surely have to be included in algorithm selection but they have to be evaluated in the context of welfare economics.