Data Dilemmas in the Information Society: Introduction and Overview - Discrimination and Privacy in the Information Society - page 21

respectively. Furthermore, Hajian and Domingo-Ferrer present metrics that can be
used to evaluate the performance of these approaches and show that
discrimination removal can be done at a minimal loss of information.
In Chapter 14, Verwer and Calders show how positive discrimination (also
known as affirmative action) can be introduced in predictive models. Three
solutions based upon so-called Bayesian classifiers are introduced. The first
technique is based on setting different thresholds for different groups. For
instance, if there are income differences between men and women in a database,
men can be given a high income label above $90,000, whereas women can be
given a high income label above $75,000. Instead of income figures, the labels
high and low income could be applied. This instantly reduces the discriminating
pattern. The second techniques focuses on learning two separate models, one for
each group. Predictions from these models are independent of the sensitive
attribute. The third and most sophisticated model is focused on discovering the
labels a dataset should have contained if it would have been discrimination-free.
These latent (or hidden) variables can be seen as attributes of which no value is
recorded in the dataset. Verwer and Calders show how decisions can be reverse
engineered by explicitly modeling discrimination.
1.4.5 Part V: Solutions in Law, Norms and the Market
Part V of this topic provides non-technological solutions to the discrimination and
privacy issues discussed in Part II. These solutions may be found in legislation,
norms and the market. Many of such solutions are discussed in other topics and
papers, such as (to name only a few) the regulation of profiling, 37 criteria for
balancing privacy concerns and the common good, 38 self-regulation of privacy, 39
organizational change and a more academic approach, 40 and valuating privacy in a
consumer market. 41 We do not discuss these suggested solutions in this topic, but
we do add a few other suggested solutions to this body of work.
In Chapter 15, Van der Sloot proposes to use minimum datasets to avoid
discrimination and privacy violations in data mining and profiling. Discrimination
and privacy are often addressed by implementing data minimization principles,
restricting collecting and processing of data. Although data minimization may
help to minimize the impact of security breaches, it has also several disadvantages.
First, the dataset may lose value when reduced to a bare minimum and, second,
the context and meaning of the data may get lost. This loss of context may cause
or aggravate privacy and discrimination issues. Therefore, Van der Sloot suggests
an opposite approach, in which minimum datasets are mandatory. This better
ensures adequate data quality and may prevent loss of context.
In Chapter 16, Finocchiaro and Ricci focus on the opposite of being profiled,
which is building one's own digital reputation. Although people have some
37 See, for instance, Bygrave, L.A. (2002).
38 Etzioni, A. (1999), p. 12/13.
39 Regan, P.M. (2002).
40 See, for instance, Posner, R.A. (2006), p. 210.
41 See, for instance, Böhme (2009) and Böhme and Koble (2007).