Examining IBM's AIF360 AI fairness platform on performance and relevance utilising ProPublica's investigation into the COMPAS recidivism dataset
Keywords
No Thumbnail Available
Authors
Issue Date
2019-06-28
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
The last decade has seen a vast increase in usage and development of Machine
Learning (ML) techniques, paired with now an increasing academical, societal,
and corporate interest in AI fairness (Gebru et al., 2018). IBM has developed
the AI fairness platform AIF360 (Bellamy et al., 2018) to facilitate the
transition of fairness research algorithms to mainstream business practise. I
show that the algorithms encased within the platform perform well on bias
detection when contrasted against ProPublica's investigation into the
COMPAS recidivism dataset, coming up with similar but weaker results. I
continue to show that the algorithms too perform well on bias mitigation as
compared to each other and an original untouched dataset when tested for
various fairness metrics. Lastly, I lay down the argument that AIF360 falls
at
on its own goal of easing the burden of speci cally ML and AI developers to
deal with questions on fairness (Bellamy et al., 2018). The absence of ways to
manage both the delayed impact of `fair' algorithms and the ambiguity of
de ning the interdisciplinary term of fairness is glaring. Furthermore, its
simpli cation of AI fairness research algorithms and metrics is not at the
request of the ML fairness expert community (Charrington, 2019), and worse
even might invoke harm due to facilitating parameter policing. I conclude that
as a whole AIF360 seems to be a positive platform for professionals in the eld
to collaborate and pool their resources together, but is neutral at best and a
danger at worst when considering its relevance and impact for mainstream
usage in business.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen