We’ve so far studied a variety of techniques and mechanisms to create programs that can classify unseen data so that it matches a “training set” well.
This works well when we assume that the classification we want to carry out should in fact directly replicate the distribution seen in the training set. At first glance, it would seem obvious that this has to be the goal. But: are you sure that the data you collected is itself what you want your ML program to do?
There are (at least) two central concerns about fairness in ML:
Technical issues. Most technical approaches to ML define a classifier that “works well” to be one that minimizes (something close to) the misclassification rate. However, there are often other concerns: what if we want to make sure our classifier doesn’t behave differently depending on the color of someone’s skin, or how old they are? How can we formulate and solve the problem under these additional design goals?
Societal issues. Most of the thinking about deploying ML overemphasizes the above technical question in excess, and ignores the context in which the ML system is going to be deployed. Who does the ML system target? Who does the ML system serve? What impact does the system have in the way people can seek redress when they’re harmed?
Although a full discussion of these issues is far beyond the scope of this class, we will engage with a number of proposals in the literature. It’s also worth noting that the literature in this topic is quite recent, which means we’ll be reading research papers in addition to textbook chapters.