Agnostic learning with unknown utilities
Abstract
Traditional learning approaches for classification implicitly assume that each mistake has the same cost. In many real-world problems though, the utility of a decision depends on the underlying context and decision . However, directly incorporating these utilities into the learning objective is often infeasible since these can be quite complex and difficult for humans to specify. We formally study this as agnostic learning with unknown utilities: given a dataset where each data point , the objective of the learner is to output a function in some class of decision functions with small excess risk. This risk measures the performance of the output predictor with respect to the best predictor in the class on the unknown underlying utility . This utility is not assumed to have any specific structure. This raises an interesting question whether learning is even possible in our setup, given that obtaining a generalizable estimate of utility might not be possible from finitely many samples. Surprisingly, we show that estimating the utilities of only the sampled points~ suffices to learn a decision function which generalizes well. We study mechanisms for eliciting information which allow a learner to estimate the utilities on the set . We introduce a family of elicitation mechanisms by generalizing comparisons, called the -comparison oracle, which enables the learner to ask for comparisons across different inputs at once. We show that the excess risk in our agnostic learning framework decreases at a rate of . This result brings out an interesting accuracy-elicitation trade-off -- as the order of the oracle increases, the comparative queries become harder to elicit from humans but allow for more accurate learning.
Cite
@article{arxiv.2104.08482,
title = {Agnostic learning with unknown utilities},
author = {Kush Bhatia and Peter L. Bartlett and Anca D. Dragan and Jacob Steinhardt},
journal= {arXiv preprint arXiv:2104.08482},
year = {2021}
}
Comments
30 pages; published as a conference paper at ITCS 2021