Machine learning is increasingly used to make decisions about people’s lives, such as whether to give someone a loan or whether to interview someone for a job. This brings with it the risk of discrimination, particularly if the data used for training the machines contains bias.
One strategy for ensuring such systems are fair is to modify the training data they learn from. Such approaches have been successful empirically, but typically lack strong theoretical fairness guarantees.
We provide mathematical definitions of some common notions of fairness, which are closely related to the concepts of equality of outcome and equality of treatment. Using these definitions, we show that it is possible to prove that a data representation is both fair and useful.
These provable properties can be used in a governance model involving a data vendor, a data user and a data regulator, where fairness and prediction accuracy are achieved separately to ensure transparency and prevent perverse incentives.
We also investigate the cost of this separation of concerns compared to a system where a single party is trusted to make predictions that are both fair and accurate. Our theoretical results motivate a practical algorithm for learning fair representations of data that can be applied in situations where both fairness and utility is required.
About the speaker
Daniel McNamara is a PhD candidate at the Research School of Computer Science of the Australian National University. He is affiliated with the Analytics Research Group at CSIRO Data61. Daniel visited the Machine Learning Department at Carnegie Mellon University in Pittsburgh as a Fulbright Postgraduate Scholar in 2016-17.