Decision Theory for Discrimination-Aware Classification

Abstract

Abstract-Social discrimination (e.g., against females) arising from data mining techniques is a growing concern worldwide. In recent years, several methods have been proposed for making classifiers learned over discriminatory data discriminationaware. However, these methods suffer from two major shortcomings: (1) They require either modifying the discriminatory data or tweaking a specific classification algorithm and (2) They are not flexible w.r.t. discrimination control and multiple sensitive attribute handling. In this paper, we present two solutions for discrimination-aware classification that neither require data modification nor classifier tweaking. Our first and second solutions exploit, respectively, the reject option of probabilistic classifier(s) and the disagreement region of general classifier ensembles to reduce discrimination. We relate both solutions with decision theory for better understanding of the process. Our experiments using real-world datasets demonstrate that our solutions outperform existing state-ofthe-art methods, especially at low discrimination which is a significant advantage. The superior performance coupled with flexible control over discrimination and easy applicability to multiple sensitive attributes makes our solutions an important step forward in practical discrimination-aware classification.

The Independence of Fairness-Aware Classifiers

“…It is easy to understand the characteristics of this alternative model because it is simple and its global optimum can be derived analyticially. We next show its connections with a two-naive-Bayes method and Kamiran et al's decision theory [7]. We empirically confirm that the degree of fairness of this alternative model is inferior to that of the two-naive-Bayes model, like the other fairness-aware classifiers.…”

“…Fairness indexes measure the degree of fairness based on observed or estimated distributions over (Y, X, S). Many types of fairness indexes have been proposed: extended lift [1], discrimination score [3], mutual information [6], [8], χ 2 -statistics [7], [9], η-neutrality [10], and a combination of statistical parity and the Lipschitz condition [11], [4]. If these fairness indexes are worse than a specified level, the corresponding decisions are considered unfair.…”

“…Data mining techniques are increasingly being used for serious determinations such as credit, insurance rates, employment applications, and so on, which are represented by Y . Needless to say, such serious determinations must guarantee fairness from both the social and legal viewpoints; that is, they must be fair and nondiscriminatory in relation to sensitive features such as gender, religion, race, ethnicity, handicaps, political convictions, and so on, which are represented by S. Discovery of discriminative treatments [13], [14], [15], [16], [1], [17], [18], [9], learning non-discriminative estimators [3], [13], [7], and transforming data for non-discriminative analysis [11], [4] have been discussed.…”

“…There are several examples of fairness-aware classification [3], [13], [8], [7], fairness-aware regression [10], and an information-neutral recommendation [5], [19]. In the case of unsupervised learning, given a data set without target values D = {x i , s i } N , the goal of fairness-aware clustering is to predict latent labels as well as to satisfy a pre-specified fairness constraint.…”

A multidisciplinary survey on discrimination analysis

“…At the time of application, instead, predictions are corrected to keep proportionality of decisions among protected and unprotected groups. Kamiran et al (2012) propose correcting predictions of probabilistic classifiers that are close to the decision boundary, given that (statistical) discrimination may occur when there is no clear feature supporting a positive or a negative decision.…”