2019/05/02 by Xueru Zhang, Zhang, Xueru, Cem Tekin +6 · 6 citations
Social Sciences · Computer Science · #Ethics and Social Impacts of AI #Mobile Crowdsensing and Crowdsourcing
paper · pdf · doi:10.48550/arxiv.1905.00569
Machine Learning (ML) models trained on data from multiple demographic groups\ncan inherit representation disparity (Hashimoto et al., 2018) that may exist in\nthe data: the model may be less favorable to groups contributing less to the\ntraining process; this in turn can degrade population retention in these groups\nover time, and exacerbate representation disparity in the long run. In this\nstudy, we seek to understand the interplay between ML decisions and the\nunderlying group representation, how they evolve in a sequential framework, and\nhow the use of fairness criteria plays a role in this process. We show that the\nrepresentation disparity can easily worsen over time under a natural user\ndynamics (arrival and departure) model when decisions are made based on a\ncommonly used objective and fairness criteria, resulting in some groups\ndiminishing entirely from the sample pool in the long run. It highlights the\nfact that fairness criteria have to be defined while taking into consideration\nthe impact of decisions on user dynamics. Toward this end, we explain how a\nproper fairness criterion can be selected based on a general user dynamics\nmodel.\n