Fairness Implications of Encoding Protected Categorical Attributes

Encoding methods for categorical data have significant fairness consequences: The paper demonstrates that how protected attributes like ethnicity or country of birth are converted into machine-readable formats (through methods like one-hot encoding or target encoding) directly affects whether resulting AI models are fair, not just accurate.

Two distinct sources of bias emerge from encoding choices: The research identifies "irreducible bias" from direct discrimination against certain groups and "reducible bias" from having too little data about underrepresented groups. Understanding which type of bias is present matters because they require different solutions.

Using protected attributes can improve fairness, but implementation details matter: While previous research showed that explicitly including protected attributes helps with fairness, this paper reveals that the technical choices made when encoding these attributes significantly influence whether that benefit actually materializes in practice.

Intersectional fairness risks increase with multiple categorical attributes: When machine learning systems encode several protected attributes together (such as both gender and ethnicity), the paper warns that standard best practices for improving performance can paradoxically create new unfairness problems for people at the intersection of multiple underrepresented groups.

Regularization techniques can mitigate encoding-related unfairness: The research shows that certain regularization methods can reduce the unfairness caused by encoding choices, suggesting that practitioners have tools available to address these fairness challenges once they're aware of them.