Introducing Positive Discrimination in Predictive Models - Discrimination and Privacy in the Information Society - page 245

Chapter 14
Introducing Positive Discrimination in
Predictive Models
Sicco Verwer and Toon Calders *
Abstract. In this chapter we give three solutions for the discrimination-aware
classification problem that are based upon Bayesian classifiers. These classifiers
model the complete probability distribution by making strong independence as-
sumptions. First we discuss the necessity of having discrimination-free classifica-
tion for probabilistic models. Then we will show three ways to adapt a Naive
Bayes classifier in order to make it discrimination-free. The first technique is
based upon setting different thresholds for the different communities. The second
technique will learn two different models for both communities, while the third
model describes how we can incorporate our belief of how discrimination was
added to the decisions in the training data as a latent variable. By explicitly model-
ing the discrimination, we can reverse engineer decisions. Since all three models
can be seen as ways to introduce positive discrimination, we end the chapter with
a reflection on positive discrimination.
14.1 Introduction
The topic of discrimination-aware data mining was first introduced in (Calders et
al., 2009; Kamiran & Calders, 2009; Pedreschi et al., 2008), and is motivated by
the observation that often training data contains unwanted dependencies between
the attributes. Given a labeled dataset and a sensitive attribute; e.g., gender, the
goal of our research is to learn a classifier for predicting the class label that does