A Fairness Approach to Mitigating Racial Bias of Credit Scoring Models by Decision Tree and the Reweighing Fairness Algorithm

Credit scoring models have been widely applied by financial institutions, Peer to Peer (P2P) lending service providers, and Buy Now Pay Later (BNPL) service providers to evaluate their customers' financial status. Therefore, it has a large impact on consumer financing activities. However, unfair evaluation may occur as the development of credit scoring models contains biased judgments (e.g., racial bias), which deteriorates users' credit access ability. Thus, we study the feasibility of mitigating racial bias in developing a credit scoring model. By using a data set provided by a P2P lending platform, LendingClub, we integrated the C5.0 decision tree algorithm and the reweighing fairness algorithm to develop credit scoring models with cost-sensitive modeling concepts. Multi-class fair credit scoring evaluation was also studied in terms of performance indices, including accuracy, average cost, and unfairness metrics. The results demonstrated that the reweighing fairness algorithm reduced the unfairness and average cost of models. In addition, combining the fairness algorithm and cost-sensitive modeling minimized the average cost of models while maintaining the functionality of the fairness algorithm.