|
Activity Number:
|
149
- Statistical Learning for Decision Support
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 8, 2022 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistical Learning and Data Science
|
|
Abstract #323654
|
|
|
Title:
|
Stochastic Ordered Empirical Risk Minimization
|
|
Author(s):
|
Ronak Mehta* and Krishna Pillutla and Vincent Roulet and Zaid Harchaoui
|
|
Companies:
|
University of Washington and University of Washington and University of Washington and University of Washington
|
|
Keywords:
|
nonsmooth optimization;
extremile;
stochastic optimization;
robust learning;
superquantile
|
|
Abstract:
|
We develop an approach to statistical learning that allows the objective to interpolate between the average loss and the worst-case loss. Our approach hinges upon the notion of the extremile, and L-statistics more generally, and leads to a learning objective that decomposes into a linear combination of order statistics of the loss. We provide a characterization of the variational properties of the objective, which we use to obtain stochastic ordered risk minimization algorithms. We present experimental results comparing the proposed algorithms to direct approaches such as stochastic subgradient and dual averaging on synthetic and real data.
|
Authors who are presenting talks have a * after their name.
