Database Reference
In-Depth Information
automated CV screening in recruitment. Given education, employment history and
qualifications (the input variables called attributes ) of an individual the task is to
decide whether this individual should be selected for an interview (the outcome
called label ). An automated classifier for such decisions can be built using historical
data examples where the relations between the attributes and labels are known.
Nowadays more and more decisions in lending, recruitment, grant or study ap-
plications are partially being automated based on models trained on historical data.
That historical data may be discriminatory 1 ; for instance, racial or gender discrim-
ination may have affected the selection of job candidates in the historical data. In
such a case classifiers trained on this discriminatory data are likely to learn the
discriminatory relation, and, as a result, they will make discriminatory predictions
when applied to new data in the future.
Training a discrimination-free model on the historical data that is discriminatory
is challenging. Removing the sensitive attribute, e.g., gender, from the training data
is not enough to prevent discrimination. If gender is related to some of the remaining
attributes, e.g, marital status, the model will capture the discriminatory decisions in-
directly. A number of techniques have been developed (Calders et al., 2009; Calders
and Verwer, 2010; Kamiran et al., 2010; Kamiran and Calders, 2010) focusing on
how to train discrimination-free classifiers over the discriminatory training data.
These techniques aim at making the probabilities of positive decision equal across
the sensitive groups, e.g., male and female. They do not take into account that some
differences in treatment may be explainable by other attributes, such as education
level. This chapter presents a methodology how to quantify and measure the ex-
plainable and non-explainable parts of discrimination and introduces classification
techniques to remove the non-explainable part only. The methodology is referred to
as conditional non-discrimination.
The studies by (Pedreschi et al., 2008, 2009; Ruggieri et al., 2010) aim at the
detection of discrimination from training data and identify the potentially discrimi-
natory classification rules. A central notion in these works on identifying discrimi-
natory rules is that of the context of the discrimination. That is, specific regions in
the data are identified in which the discrimination is particularly high. These works
focus also on the case where the discriminatory attribute is not present in the dataset
and background knowledge for the identification of discriminatory guidelines has to
be used. However, we assume that the discriminatory data is given but the discrimi-
nation should be avoided in future predictions. Our chapter discusses the next steps
after detecting discrimination.
The chapter is organized as follows. In Section 8.2 we discuss the problem of
explainable and illegal discrimination in classifier design. Section 8.3 analyzes the
concept of explainable and non-explainable discrimination, and instructs how to
measure the explainable part of the discrimination. In Section 8.4 we present the
local modeling techniques for removing illegal discrimination and experimentally
illustrate their performance. Section 8.5 concludes the chapter.
1
Discrimination is the prejudicial treatment of an individual based on their membership in
a certain group or category.
Search WWH ::




Custom Search