<p>While many studies focus on binary bias mitigation methods, there is a lack of research on multi-class settings. This study adapts well-established pre-processing binary bias mitigation techniques for multi-class configurations to reduce bias with minimal accuracy loss. We adapt six binary bias mitigation methods for multi-class settings and compare them to baseline model with 3 to 5 class, while also evaluating the performance of binary methods in a binary configuration, assessing both accuracy and discrimination by class. In parallel, we compare wage discrimination patterns between Brazil and New York using descriptive and inferential statistics, wage decomposition and Ordinary Least Squares (OLS) regression. Our results demonstrate, with statistical significance, that suppressing data reduces bias by up to 75% with over 10% accuracy loss, and adaptations of the reweighing method allow for up to 55% reduction with minimal loss relative to baseline methods. Furthermore, similar patterns of wage discrimination are observed in both the New York and Brazilian datasets, with nuanced interactions between sensitive attributes. Binary bias mitigation methods can be adapted for multi-class classification, yielding positive results by reducing bias with minimal accuracy loss.</p>