A Fair Classifier Using Kernel Density Estimation

Lecture 2: A fair classifier using kernel density estimation Recap At the beginning of the last lecture, I mentioned that trustworthy AI is a new and trending topic that we are going to touch upon in this tutorial. And I told you that among several aspects that can represent trustworthy AI, the following two are of this tutorial's focus: (i) fairness (targeting unbiased decisions among different demographics and/or individuals); and (ii) robustness (pursuing an interested model being robust to data poisoning). In particular, we aimed to explore the two issues in the context of classifiers with a particular emphasis on one prominent fairness concept, called group fairness, aiming for irrelevancy of predictions to sensitive attributes such as race, gender, age and religion. We then introduced two fairness measures that quantify the degree of group fairness. The first is DDP which promotes the independence between sensitive attribute Z and prediction Ỹ (made in hard decision):