讲座:Metrizing Fairness 发布时间:2025-05-07

  • 活动时间:
  • 活动地址:
  • 主讲人:

题 目:Metrizing Fairness

嘉 宾:Daniel Kuhn, 教授,洛桑联邦理工学院

主持人:许欢 教授 上海交通大学安泰经济与管理学院

时 间:2025年5月13日(周二)15:00-16:30

地 点:Zoom会议(校内师生如需会议号和密码,请发邮件至yuxin.su@sjtu.edu.cn获取),线下会场:安泰楼B207室

内容简介:

We study supervised learning problems that have significant effects on individuals from two demographic groups, and we seek predictors that are fair with respect to a group fairness criterion such as statistical parity (SP). A predictor is SP-fair if the distributions of predictions within the two groups are close in Kolmogorov distance, and fairness is achieved by penalizing the dissimilarity of these two distributions in the objective function of the learning problem. In this paper, we identify conditions under which hard SP constraints are guaranteed to improve predictive accuracy. We also showcase conceptual and computational benefits of measuring unfairness with integral probability metrics (IPMs) other than the Kolmogorov distance. Conceptually, we show that the generator of any IPM can be interpreted as a family of utility functions and that unfairness with respect to this IPM arises if individuals in the two demographic groups have diverging expected utilities. We also prove that the unfairness-regularized prediction loss admits unbiased gradient estimators, which are constructed from random mini-batches of training samples, if unfairness is measured by the squared L2-distance or by a squared maximum mean discrepancy. In this case, the fair learning problem is susceptible to efficient stochastic gradient descent (SGD) algorithms. Numerical experiments on synthetic and real data show that these SGD algorithms outperform state-of-the-art methods for fair learning in that they achieve superior accuracy-unfairness trade-offs -sometimes orders of magnitude faster.

演讲人简介:

Daniel Kuhn is a Professor of Operations Research in the College of Management of Technology at EPFL, where he holds the Chair of Risk Analytics and Optimization. His research interests revolve around optimization under uncertainty. Before joining EPFL, Daniel Kuhn was a faculty member in the Department of Computing at Imperial College London and a postdoctoral researcher in the Department of Management Science and Engineering at Stanford University. He holds a PhD degree in Economics from the University of St. Gallen and an MSc degree in Theoretical Physics from ETH Zurich. He is an INFORMS fellow and the recipient of several research and teaching prizes including the Friedrich Wilhelm Bessel Research Award by the Alexander von Humboldt Foundation as well as the Frederick W. Lanchester Prize and the Farkas Prize by INFORMS. He is the editor-in-chief of Mathematical Programming.

 

 

欢迎广大师生参加!