Speaker Jason Hartline
Affiliation Northwestern University
Host Nicole Immorlica
Date recorded 19 March 2014
The promise of data science is that system data can be analyzed and its understanding can be used to improve the system (i.e., to obtain good outcomes). For this promise to be realized, the necessary understanding must be inferable from the data. Whether or not this understanding is inferable often depends on the system itself. Therefore, the system needs to be designed to both obtain good outcomes and to admit good inference. This talk will explore this issue in a mechanism design context where the designer would like use past bid data to adapt an auction mechanism to optimize revenue. Data analysis is necessary for revenue optimization in auctions, but revenue optimization is at odds with good inference. The revenue-optimal auction for selling an item is typically parameterized by a reserve price, and the appropriate reserve price depends on how much the bidders are willing to pay. This willingness to pay could be potentially be learned by inference, but a reserve price precludes learning anything about willingness-to-pay of bidders who are not willing to pay the reserve price. The auctioneer could never learn that lowering the reserve price would give a higher revenue (even if it would). To address this impossibility, the auctioneer could sacrifice revenue-optimality in the initial auction to obtain better inference properties so that the auction's parameters can be adapted to changing preferences in the future. In this talk, I will develop a theory for optimal auction design subject to good inference.
©2014 Microsoft Corporation. All rights reserved.