Mitigating bias in machine learning : towards discrimination-free classification via fairness aware bagging
Loading...
Date
Authors
Παπαγιαννόπουλος, Παναγιώτης
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
It is considered a well known truth that Artificial Intelligence (AI) is changing and will
change the world that we live in. Many have predicted that it will bring the second biggest
change, after the electronically interconnected world (the internet), by 2030. That positive
impact can only be enhanced by combining it with other technological advances, such as
Big Data and the growing adoption of automation and robotics.
There is no doubt that Artificial Intelligence has the insight to make decisions and it
already provide us with countless benefits. We must recognize, however, that the algorithms
used, as well as the people who have created them, are not perfect for consistently
providing us with perfect answers and calculations. Since people are inherently biased
in their decision-making - making consciously and unconsciously biased decisions in their
everyday lives - their algorithms using Artificial Intelligence are led to the same fate. So,
while these algorithms push us to make leaps and bounds in tackling some of the world’s
most demanding problems, they can also provoke distrust and promote mass discrimination
on a mass scale with little transparency or accountability, reducing people’s chances
of participating equally not only in economies but also in society.
The goal of this thesis is to list at their full extent, the problems related to the biases
of algorithms, present the most recent approaches that aim to make Discrimination-free
algorithms and finally present a novel method based on the existing ones which will make
Discrimination-Free Classifiers by achieving good Fairness score without sacrificing its
Accuracy. This novel method is called Fairness Aware Bagging (FAB).
Description
Keywords
Bias in machine learning, Discrimination in AI, Fairness, Transparency, Accountability, FAT