Think Twice Before Enforcing A Notion: A Reflection and a Case Study on Fairness in Machine Learning

The thesis presents a reflection and a case study of fairness notions in machine learning. I review commonly used fairness notions and reflect on the subtleties with respect to the role played by causality in fairness analysis. Then focusing on the Equalized Odds notion of fairness, I consider the theoretical attainability of Equalized Odds and, furthermore, if it is attainable, the optimality of the prediction performance under various settings. In particular, for prediction performed by a deterministic function of input features, I give the conditions under which Equalized Odds can hold true; if the stochastic prediction is acceptable, I show that under mild assumptions, fair predictors can always be derived. For classification tasks, I further prove that compared to enforcing fairness by post-processing, one can further benefit from exploiting all available features during training and get potentially better prediction performance while remaining fair.