Implementing Fair Regression In The Real World

In a business context where an unconstrained real-world application were The potential risk of machine learning algorithms to unintentionally to be replaced with a fairer one, such extreme discrepancies embed and reproduce bias and therefore discriminating would not be viable because individuals who were substantially various sub populations in high-stakes decisionmaking negatively impacted would probably not accept applications has given rise to the new research the change and switch to a competitor. Based on our findings, field of fair machine learning (Kamiran and Calders 2009; we therefore propose algorithmic post-processing procedures Corbett-Davies et al. 2018; Barocas, Hardt, and Narayanan to adjust for unwanted, extreme discrepancies between 2019). Plenty of quantitative measures of fairness have been unconstrained and fair methods in order to enable a proposed (Dwork et al. 2011; Hardt, Price, and Srebro 2016; smooth transition from an "unfair" to a fairer model. Chouldechova 2017; Berk et al. 2018) which opened up The main contributions of this paper are: the way for three types of algorithms that seek to satisfy them: First, the pre-processing approach which modifies - We empirically examine the evolution of fair regression the data representation prior to using classical algorithms outputs compared to unconstrained predictors and demonstrate (Kamiran and Calders 2012; Zemel et al. 2013). Second, that some variations on the individual level may be the in-processing approach which intervenes during the unacceptable in practice. To the best of our knowledge we learning phase by adding a fairness constraint to the optimization offer the first investigation of this kind; objective (Kamishima et al. 2012; Zafar et al. - We propose a range of post-processing algorithms to mitigate 2017; Zhang, Lemoine, and Mitchell 2018). Third, the postprocessing this effect and therefore provide mechanisms to approach which adjusts the outputs of classical implement fair regression in practice.