Mitigating fairness-related harms by AI/ML models is an important topic of research. The simplest formulation for theoretical study is of binary classification under fairness constraints such as demographic parity (i.e., equal acceptance rates) and equal opportunity (i.e., equal False Negative Rates) when group & class-conditioned distributions are simple parametric families. The same extends to generative models for downstream classification task. Bayes Optimal Classifier (BOC) minimizes the error rate for a given data distribution. We study its fairness-constrained analog, fair BOC, and ask two questions (a) What is a fair data distribution and when/how can we recover fair classifiers/distributions even from biased data? (b) When is the error rate of fair BOC stable under distribution shifts? Both reveal an important role of probability and randomness in fair AI/ML.
Zoom Link: https://us02web.zoom.us/j/88670406480
Meeting ID: 886 7040 6480