Incorporating Intersectional Fairness Into Machine Learning

Elyjah Bassford

Mentor: Dr. James Foulds

Department of Information Systems, UMBC

With the prevalence of machine learning algorithms in everyday life, such as in social media platforms and job recommendations, unbiased outcomes necessitate proper definitions of fairness. Predictive policing, using algorithms to predict crime locations and likelihood of further offenses, among other actions, particularly requires a check against bias. In this work, we utilize differential fairness, a definition of fairness that incorporates the modern intersectional fairness principle, in producing a machine learning algorithm to accurately predict crime locations. We propose exploiting the structure of convolutional neural networks to maximize tasks performed by the algorithm. This definition of fairness will greatly exceed standard accuracy, among other types of machine learning algorithms, compared to algorithms using definitions such as infra-marginality which does not compensate for societal bias and often resulting in the targeting of certain demographics.