The paper addresses the problem of identifying and removing disparate impact in algorithms, a concept that is legally defined in the U.S. as unintentional discrimination based on protected attributes (e.g., race, gender). The authors propose a method to certify that an algorithm does not have disparate impact and to transform data to remove it. They link disparate impact to the balanced error rate (BER) and develop a regression algorithm to minimize BER, which is then used to certify the absence of disparate impact. They also describe methods to transform data to make it unbiased while preserving classification accuracy. Empirical results show that their approach effectively certifies the absence of disparate impact and preserves relevant information in the data. The paper includes a detailed analysis of the theoretical underpinnings, computational fairness, and the tradeoff between fairness and utility in partial repairs.The paper addresses the problem of identifying and removing disparate impact in algorithms, a concept that is legally defined in the U.S. as unintentional discrimination based on protected attributes (e.g., race, gender). The authors propose a method to certify that an algorithm does not have disparate impact and to transform data to remove it. They link disparate impact to the balanced error rate (BER) and develop a regression algorithm to minimize BER, which is then used to certify the absence of disparate impact. They also describe methods to transform data to make it unbiased while preserving classification accuracy. Empirical results show that their approach effectively certifies the absence of disparate impact and preserves relevant information in the data. The paper includes a detailed analysis of the theoretical underpinnings, computational fairness, and the tradeoff between fairness and utility in partial repairs.