IEEE Technology and Society Magazine - June 2021 - 75

between individuals is fairness through awareness
[17], which states that, for fairness to hold, the distance
between the distributions of outputs for individuals
should at most be the distance between the
two individuals as estimated by means of a similarity
metric. The complexity in using this metric consists
in accurately defining a similarity measure that correctly
represents the complexity of the situation in
question, which is often an impossible task to generalize.
Moreover, the similarity measure between
individuals can suffer from the implicit biases of the
expert, resulting in a biased similarity estimator.
Finally, definitions based on causal reasoning
assume bias can be attested by means of a directed
causal graph. In the graph, attributes are presented
as nodes joined by edges which, by means of equations,
represent the relations between attributes [10].
By exploring the graph, the effects that the different
protected attributes have on the algorithm's output
can be assessed and analyzed. Causal fairness
approaches are limited by the assumption that a valid
causal graph able to describe the problem can be
constructed, which is not always feasible due to the
sometimes unknown and complex relations between
attributes and the impact they have on the output.
Attesting and addressing
discrimination
The first step explored in the related literature to
identify discriminatory outputs is determining the
groups whose algorithmic outputs are going to be
compared. Technical approaches to select the subpopulations
of interest vary, either: 1) they consider
subpopulations as already defined [9], [18] or 2)
they are selected by means of a heuristic that aggregates
individuals that share one or more protected or
proxy attributes (protected groups), as in FairTest's
framework1 for detecting biases in data sets. Protected
attributes are encoded in legislation (see the " Legal
perspective " section) and usually include attributes
such as sex, gender, and ethnicity, while proxy attributes
are attributes strongly correlated with protected
attributes, e.g., weightlifting ability (strongly correlated
with gender). However, the process of selecting
individuals or groups based on these attributes is nontrivial
since groups often result from the intersection
of multiple protected and proxy attributes (see the
" Social perspective " section).
1https://github.com/columbia/fairtest
Once the protected and the potentially advantaged
groups have been selected, implementations
apply different bias metrics (see the " Bias metrics "
section) to compare and identify relevant differences
in the algorithm's outcomes for the different
groups. If these differences are a consequence of
protected attributes, it is likely that the algorithm's
decision can be considered discriminatory.
To alleviate the contextual problem of whether an
algorithmic outcome may form a case of discrimination,
approaches often incorporate explanatory attributes:
attributes, such as gender or age, on which in
specific contexts is deemed acceptable to differentiate,
even if this leads to apparent discrimination on
protected attributes [18]. Some relevant approaches
are the open-source IBM AI Fairness 360 toolkit,2
which contains techniques developed by IBM and
the research community to help detect and mitigate
bias in machine learning models throughout the
AI application lifecycle, and Google's Whatif-tool,3
which offers an interactive visual interface that allows
researchers to investigate model performances for
a range of features in the data set and optimization
strategies.
Despite these efforts in parameterizing context
uncertainty in technical implementations, the interpretive
dimension that separates bias and discrimination
remains a challenge. As a response, some approaches
base their implementations on various antidiscrimination
laws that focus on the relationships between protected
attributes and decision outcomes. For instance,
the U.S. fourth-fifth court rule and the Castaneda rule are
used as a general, and often arguably adequate, prima
facie evidence of discrimination-see the " Legal perspective "
section for more details on these rules.
Approaches that intervene on problematic biases
focus on: 1) removing protected attributes from the
data, as an attempt to impede the algorithm from
using these protected attributes to make discriminatory
decisions (fairness through blindness [12],
[17]), or on 2) debiasing algorithms' outputs [19].
An issue here is that removing protected attributes
from the input data often results in a significant
loss of accuracy in the algorithm [17]. Moreover,
excluded attributes can often be correlated with
proxy attributes that remain in the data set, meaning
bias may still be present (i.e., certain residential
2https://github.com/IBM/AIF360
3https://pair-code.github.io/what-if-tool/
June 2021
75
IEEE Technology and Society Magazine - June 2021

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