Abstract: Data-quality issues often compound bias in machine learning, since majority groups are usually well-studied, with copious high-quality data available, whereas marginalized or minority groups are understudied, and available data lack in quality. We operate in a supervised learning setting, i.e., we learn a function h : X - Y, wherein only partial information is available on protected group membership Z. In particular, each datapoint corresponds to a triplet (x, y, z) ? (X x Y x Z), where Z = {1, 2, ... , g} is a finite space of g protected groups. Given m training points, we observe features x and labels y, but not group identities z. Instead, we are given a feasible set Z of labelings. The task is then to perform group-dependent fair (supervised) learning, with rigorous analytical guarantees.
We show that learning approximately minimax-optimal egalitarian or utilitarian malfare models in this setting is both statistically and computationally tractable. In particular, our generalization bounds depend on how sharply the unknown group-membership labels are constrained, and thus degrade gracefully as less and less partial information about group membership is available. We also discuss methods by which to statistically constrain the feasible set Z of group membership and the fairness implications of generating such constraints.
Bio: Known to some as The Count of Monte Carlo, Cyrus Cousins studies all manner of problems involving sampling, randomization, and learning in data science and beyond, with a particular interest in uniform convergence theory and rigorous treatment of fair machine learning. He is currently serving as a visiting assistant professor at Brown University, where he also completed his doctoral studies under the tutelage of the great Eli Upfal. Cyrus was awarded the Joukowsky Outstanding Dissertation Prize (in the physical sciences) for his dissertation, "Bounds and Applications of Concentration of Measure in Fair Machine Learning and Data Science," and before arriving at Brown University, he earned his undergraduate degree in computer science, mathematics, and biology from Tufts University.
The CICS Theory Seminar is free and open to the public. If you are interested in giving a talk, please email Cameron Musco or Rik Sengupta. Note that in addition to being a public lecture series, this is also a one-credit graduate seminar (CompSci 891M) that can be taken repeatedly for credit.