Skip to Main Content
Regulations worldwide ban discrimination on many factors, including gender, race, age, etc. This poses a problem for data mining, as learning from historical data containing discriminatory decisions may perpetuate discrimination, even if protected attributes are not used. We focus on discrimination prevention for classification. We introduce a new training set correction approach to handle discriminatory decision policies. Previous training set correction approaches are policy-neutral, our approach specifically targets decision policies evidencing discrimination. The goal is to target specific evidence of discrimination, and thus reduce discrimination with little impact on classification accuracy.