ADVERSARIAL LEARNING METHODS FOR FAIR MACHINE LEARNING
Keywords
Loading...
Authors
Issue Date
2019-08-02
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
As machine learning systems grow to take up ever more space in our everyday lives, ensuring
that they are aligned with human values became an important topic. One issue of automatic
systems which learn from large amount of often human-labeled data is that biases and
discriminatory practices which are incorporated in the data are taken over or amplified during
the learning process. Fair machine learning has emerged as a field of research, dedicated to
discover and study methods to alleviate such biases. In this work, the method of adversarial
debiasing was employed using a neural network model on two well known datasets. A novel
weighting term from another field of machine learning was introduced to this domain and
experimentally validated. Results show no consistent benefit from this weighting and further
experimentation needs to be done to deploy its full potential. Furthermore, the methodology
was extended to work for more than one feature which is commonly discriminated against.
First findings indicate that by using one of the proposed extensions, two sensitive variables
can successfully be protected during classification.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen