IEEE Technology and Society Magazine - June 2021 - 74

Feature
Procedural approaches
focus on identifying
biases in the algorithmic " logic. " Such ante hoc
interventions are hard to implement for two main
reasons: 1) AI algorithms are often sophisticated and
complex since, in addition to being trained on huge
data sets, they usually make use of unsupervised
learning structures that might prove difficult to trace
and understand (e.g., neural networks) and 2) the
source code of the algorithm is rarely available. Procedural
approaches will become more beneficial
with further progress in explainable AI [6].
Being able to understand the process behind
an algorithmic discriminatory decision can help us
understand possible problems in the algorithm's
code and behavior, and thus act accordingly toward
the creation of nondiscriminatory algorithms. As
such, current literature on nondiscriminatory AI
promotes the introduction of explanations into the
model itself, e.g., through inherently interpretable
models such as decision trees, association rules,
causal reasoning, or counterfactual explanations
which provide coarse approximations of how a system
behaves by explaining the weights and relationships
between variables in (a segment of) a model
[7]-[11]. Note, however, that attesting that an algorithmic
process is free from biases does not ensure
a nondiscriminatory algorithmic output, since discrimination
can arise as a consequence of biases in
training or in usage [12].
While procedural approaches attend to the algorithmic
process, relational
approaches measure
biases in the data set and the algorithmic output.
Such approaches are popular in the literature, as they
do not require insights into the algorithmic process.
Besides evaluating biases in the data itself, where it is
available (e.g., by looking at statistical parity), implementations
can compare the algorithmic outcomes
obtained by two different subpopulations in the data
set [13], or make use of counterfactual or contrastive
explanations [8], [11], [14], which have shown
promising results in aiding the provision of interpretable
models and make the decisions of inscrutable
systems intelligible to developers and users, by asking
questions such as " What if X instead of Y? "
Bias, here, is only located at testing time. One
example is the post-hoc approach of local interpretable
model-agnostic explanations (LIME), which
makes use of adversarial learning to generate counterfactual
explanations [6]. Other approaches evaluate
the correlation between algorithmic inputs and
74
biased outputs, to identify those features that may
lead to biased actions that affect protected subpopulations
[15]. Since implementations often ignore the
context in which the algorithm will be deployed, the
decision of whether a biased output results in a case
of discrimination is often left to the user to assess [9].
Bias metrics
The metrics for measuring bias can be organized
in three different categories: 1) statistical measures;
2) similarity-based measures; and 3) causal reasoning.
While reviews such as [16] offer an extensive
description of some of these metrics, we will discuss
the intuition behind the most common types of metrics
used in the literature below.
Statistical measures to attest biases represent the
most intuitive notion of bias, and focus on exploring
the relationships or associations between the
algorithm's predicted outcome for the different
(input) demographic distributions of subjects, and
the actual outcome that is achieved. These measures
include, first, group fairness (also named statistical
parity), which requires that an equal quantity
of each group of distinct individuals should receive
each possible algorithmic outcome. For instance, if
four out of five applicants of the advantaged group
were given a mortgage, the same ratio of applicants
from the protected group should obtain the mortgage
as well. Second, predictive parity is satisfied if
both protected and unprotected groups have equal
positive predictive value-that is, the probability of
an individual to be correctly classified as belonging
to the positive class. Finally, the principle of well
calibration states that the probability estimates provided
by the decision-making algorithm should be
properly adjusted with the real values. Despite the
popularity of statistical metrics, it has been shown
that statistical definitions are insufficient to estimate
the absence of biases in algorithmic outcomes, as
they often assume the availability of verified outcomes
necessary to estimate them, and often ignore
other attributes of the classified subject than the sensitiveones
[17].
Similarity measures, on the other hand, focus
on defining a similarity value between individuals.
Causal discrimination is an example of such measures,
stating that a classifier is not biased if it produces
the same classification for any two subjects
with the same nonprotected attributes. A more
complex bias metric based on a similarity measure
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2021
