Observational fairness

Equality is often seen as a purely qualitative issue, and as such, it's often dismissed by quantitative-minded modelers. As this section will show, equality can be seen from a quantitative perspective, too. Consider a classifier, c, with input X, some sensitive input, A, a target, Y and output C. Usually, we would denote the classifier output as , but for readability, we follow CS 294 and name it C.

Let's say that our classifier is being used to decide who gets a loan. When would we consider this classifier to be fair and free of bias? To answer this question, picture two demographics, group A and B, both loan applicants. Given a credit score, our classifier must find a cutoff point. Let's look at the distribution of applicants in this graph:

Note

Note: The data for this example is synthetic; you can find the Excel file used for these calculations in the GitHub repository of this book, https://github.com/PacktPublishing/Machine-Learning-for-Finance/blob/master/9.1_parity...