Exploiting reject option in classification for social discrimination control

Abstract

Social discrimination is said to occur when an unfavorable decision for an individual is influenced by her membership to certain protected groups such as females and minority ethnic groups. Such discriminatory decisions often exist in historical data. Despite recent works in discrimination-aware data mining, there remains the need for robust, yet easily usable, methods for discrimination control. In this paper, we utilize reject option in classification, a general decision theoretic framework for handling instances whose labels are uncertain, for modeling and controlling discriminatory decisions. Specifically, this framework permits a formal treatment of the intuition that instances close to the decision boundary are more likely to be discriminated in a dataset. Based on this framework, we present three different solutions for discrimination-aware classification. The first solution invokes probabilistic rejection in single or multiple probabilistic classifiers while the second solution relies upon ensemble rejection in classifier ensembles. The third solution integrates one of the first two solutions with situation testing which is a procedure commonly used in the court of law. All solutions are easy to use and provide strong justifications for the * Corresponding author A C C E P T E D M A N U S C R I P T decisions. We evaluate our solutions extensively on four real-world datasets and compare their performances with previously proposed discrimination-aware classifiers. The results demonstrate the superiority of our solutions in terms of both performance and flexibility of applicability. In particular, our solutions are effective at removing illegal discrimination from the predictions.

Fairea: a model behaviour mutation approach to benchmarking bias mitigation methods

“…Two post-processing methods that optimise for EOD, are Equalised Odds (EO) [28,51] and Calibrated Equalised Odds [51]. Other post-processing approaches include the modification of the probability of positive decisions for Naive Bayes (NB) [14], leaf relabelling for Decision Trees (DT) [37], and further investigation of uncertain labels [39].…”

“…To compare the fairness-accuracy trade-off achieved by bias mitigation methods, practitioners either observe the fairness and accuracy changes in separate graphs, or visualise them in a 2dimensional graph (one dimension is accuracy, the other dimension is fairness) [12, 14-16, 24, 36, 37, 39, 41, 42, 51, 60]. The proposed mitigation methods are often compared with previous methods [16, 17, 37-41, 51, 61, 69], different configurations [12,14,24,[38][39][40], the original non-optimised classifier [12,15,36,61,62], or a classifier trained without using protected attributes [12,15,36,69]. [11] Interact Comput Evaluation of problem-solving software for gender-inclusiveness.…”

“…ML fairness, an important non-functional testing property of ML software [66], has been widely studied in the past few years in both software engineering [10,17,31,64,66] and machine learning literature [5,14,38,40]. Some approaches adapt the training data to reduce data bias (i.e., pre-processing) [15,17,24,36], some create classification models that consider fairness during the training process (i.e., in-processing) [5,16,41,60,62], others apply changes to the model prediction outcomes to reduce bias (i.e., postprocessing) [14,28,37,39,51].…”

“…Previous work presented the trade-offs in a qualitative manner. They either report and analyse the bias mitigation effectiveness by plotting the accuracy and fairness for a visual comparison [36,39,51], or display accuracy and fairness separately [17,40,61,69] (in tables or bar charts). As far as we know, there is no trade-off baseline, nor is there any quantitative approach that can automatically evaluate and compare the fairness-accuracy trade-offs of software bias mitigation methods.…”

Fairway: a way to build fair ML software

“…• Pre-processing algorithms: In this approach, before classification, data is pre-processed in such a way that discrimination or bias is reduced. Kamiran [38] which gives favorable outcomes to unprivileged groups and unfavorable outcomes to privileged groups within a confidence band around the decision boundary with the highest uncertainty.…”