An Experiment on Leveraging SHAP Values to Investigate Racial Bias

We design a series of experiments on credit scoring and employ SHAP values to demonstrate that the use of location information may introduce racial biases. The analysis relies on race statistics collected from Brazilian Institute of Geography and Statistics (IBGE) and on fully anonymized credit information. The present work helps to discuss how to track racial biases when protected attributes are not available and points in interesting directions towards the development of procedures to yield more ethical credit scoring models in the Brazilian context.