Transparency in Data Mining: From Theory to Practice - Discrimination and Privacy in the Information Society - page 293

forced to merely state that the individual was singled out based on the algorithm,
which was structured on the basis of previous findings. 7
A policy decision mandating interpretable results calls upon analysts to work
through the statistical outputs received, understand their meaning and articulate
them clearly. In doing so, analysts note the correlations between higher risks and
personal factors (such as height, age, specific credit or purchasing history). With
this information, the analyst sets up profiles based on these findings, while
defining their parameters, and applies them to future events. When seeking
correlations, analysts might choose to ignore findings which seem ridiculous or
cannot be explained by an intuitive causation model. Thus, interpretability could
be considered as an important step to assure quality and precision, and that the
results are not merely anecdotal. The analyst could also provide a response to
external inquiries as to what initiated special treatment of an event or individual.
The flip side is that interpretability calls for models which are less complex and
therefore less accurate (Martens & Provost, 2011).
Interpretability also allows the analyst to go beyond correlation and search for a
theory that could uncover causation. For instance, one way, cash-only airline
tickets could (in theory) be casually linked to terrorists planning to ignite
explosives on an aircraft. Constructing a theory of causation linking these two
dynamics is relatively simple (although not necessarily true). Other correlations
might call for more elaborate theories of causations. Validating such theories will
call for additional study both of fact patterns and possibly in the field - all in an
attempt to reveal the forms of causation in play. Therefore, requiring a theory of
causation to be set in place prior to taking action based on correlations would
further assure the precision of the process. On the other hand, requiring causation
theories might potentially slow down and encumber the efficiency of the entire
process (and might even be an impossible task). In summary, policy decisions
mandating interpretability and causation are subtle, but will have a substantial
impact on the prospect of transparency throughout the process.
17.3 The Nature of Transparency in Predictive Modeling:
Working through the Information Flow
A call for transparency evolves when considering predictive data mining and its
outcomes. Yet transparency can refer to a variety of segments throughout the
prediction modeling process. Assuring transparency at every segment generates
specific forms of costs and balances, and is derived from a different set of laws
and justifications. In some instances, transparency might merely require uploading
7 This is mostly the case when more advanced tools of data mining are applied, such as
decision tree learning. Since these tools generate specific concerns of their own, they will
not be further addressed here. For a discussion of such instances that at times involved
tens of thousands of factors, see David Martens & Foster Provost, Explaining Documents'
Classification , NYU - Stern School of Business, Working Paper CeDER-11-01,
http://archive.nyu.edu/handle/2451/29918.