Informs Annual Meeting 2017

TB09

INFORMS Houston – 2017

TB09

2 - Guided Undersampling Classification for Automated Radiation Therapy Quality Assurance of Prostate Cancer Treatment William Eric Brown, Texas A&M.University, 3131 TAMU, College Station, TX, 77843-3131, United States, webrown@tamu.edu, Kisuk Sung, Dionne Aleman, Erick Moreno-Centeno, Thomas Purdie, Chris J. McIntosh Review of radiation therapy (RT) treatment plans requires ample care and human capital; failure to identify errors can cause sizable patient harm. We cast this issue as imbalanced-data classification and address it with a novel modularized guided undersampling method. A sequence of accepted and erroneous RT plans for low to intermediate risk prostate cancer patients was used for testing. While upholding clinically acceptable false-positive rates (FPR), we identify 100% of flawed plans. Remarkably, our framework correctly identifies error types it was not trained on with an FPR of only 18%. 3 - Optimizing the Emergency Call Taking Process with Machine Learning Methods. Nikiah Nudell, Dakota State University, 2800 7th St N, St. Cloud, MN, 56303, United States, nnudell@paramedicfoundation.org Over 7 billion emergency calls are made worldwide to emergency numbers such as 911 each year. For cardiac arrest cases, delays cost lives and the currently used decision tools only achieve about 85% accuracy. This study proposes the first work to optimize the emergency call taking process based on patient outcomes using an adaptive hybrid machine learning method. Our method could be implemented in emergency call centers to improve the accuracy and speed of incident detection for up to 50% improvement in survival benefit. These self- optimizing methods could be deployed across many industries from military to retail, to manufacturing, to customer service upstream of decision support systems. 4 - A Hybridized Machine Learning Framework for Health Risk Ranking Doug Popken, Senior Vice President, NextHealth Technologies, Denver, CO, United States, doug.popken@nexthealthtechnologies.com, Jeremy Schendel We have deployed a hybrid predictive modeling system in a production environment to predict different forms of medical utilization. Predictions are used for risk stratification and subsequent targeting of behavior changing interventions. Challenges and ongoing research directions are presented. 322B OR and IS in Health: Never the Twain Shall Meet? Sponsored: Health Applications Sponsored Session Chair: Sally C. Brailsford, University of Southampton, Southampton, S017 1BJ, United Kingdom, s.c.brailsford@soton.ac.uk 1 - Panelist Cynthia M.Lerouge, University of South Florida, 1727 Carroll Street, Saint Louis, MO, 63104-3305, United States, lerouge@tampabay.rr.com 2 - Panelist Fay Cobb Payton, North Carolina State University, COM, Box 7229, Raleigh, NC, 27695, United States, fay_payton@ncsu.edu 3 - Panelist Paul Harper, Cardiff University, Cardiff, United Kingdom, harper@cardiff.ac.uk 4 - Operations Research and Information Systems: Never the Twain Shall Meet? TB08

330A Online Advertising Markets Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session

Chair: Gabriel Weintraub, Stanford Graduate School of Business, Stanford, CA, 94304, United States, gweintra@stanford.edu 1 - Budget Management Strategies in Repeated Auctions Anthony Kim, Stanford University, 39 Angell Court Apt. 106, Stanford, CA, 94305, United States, tonyekim@stanford.edu, Santiago Balseiro, Mohammad Mahdian, Vahab Mirrokni One of the most important features advertising platforms often provide is budget management, which allows advertisers to control their cumulative expenditures. We study the “system equilibria” of a range of budget management strategies in terms of the seller’s profit, buyers’ utility and overall welfare. We show these methods admit a system equilibrium in a rather general asymmetric setting and prove dominance relations between them when buyers are symmetric. We also empirically compare the system equilibria of these strategies using real ad auction data. 2 - Price Discrimination with Ex-post Participation Constraints Francisco Castro, Columbia University, New York, NY, United States, fcastro19@gsb.columbia.edu, Dirk Bergemann, Gabriel Weintraub In online display advertising markets, publishers typically use a series of second- price auctions to “waterfall” a given impression, implicitly imposing different priorities and reserve prices over buyers. In addition, business constraints do not allow sellers to charge up-front fees. Thus motivated, we study a sequential screening problem with ex-post participation constraints. Our results shed light on the effectiveness of the series of waterfall auctions as a price discrimination device and on what type of improvements could be implemented in practice. 3 - Boosted Second-price Auctions, Simple Mechanisms to Capture Heterogeneity Negin Golrezaei, University of Southern California, 812 S Marengo Avenue, Unit 6, Pasadena, CA, 91106, United States, golrezae@usc.edu, Max Lin, Vahab Mirrokni, Hamid Nazerzadeh We propose a new auction format called boosted second-price auctions that can yield more revenue than widely used second-price auctions. In boosted second- price auctions, for each advertiser, a boosted bid is calculated, where the boosted bid is a linear function of submitted bids. Then, an advertiser with the highest boosted bid wins. We provide theoretical justification as to how new boosted second-price auctions increase revenue over second-price auctions when advertisers are heterogeneous. We further provide practical guidelines to determine boosts. We validate our guidelines theoretically and empirically using Google Ad Exchange data. 4 - Multiplicative Pacing Equilibria in Auction Markets Christian Kroer, Carnegie Mellon University, Computer Science Department, 5000 Forbes Avenue, Pittsburgh, PA, 15213, United States, ckroer@cs.cmu.edu, Vincent Conitzer, Eric Sodomka, Nicolas Stier-Moses Budgets play a significant role in auction markets implemented by Internet companies. To maximize the value provided to auction participants, spending is smoothed across auctions so budgets are used for the best opportunities. We consider a smoothing procedure that relies on pacing multipliers: for each bidder, the platform applies a factor between 0 and 1 that uniformly scales the bids across all auctions. We introduce the notion of pacing equilibrium, and prove that one or more equilibria always exist. Although computing a social welfare or revenue maximizing pacing equilibrium is NP-hard, we develop an integer program that can be used to characterize pacing equilibria and optimize among them.

Sally C. Brailsford, University of Southampton, School of Management, Southampton, S017 1BJ, United Kingdom, s.c.brailsford@soton.ac.uk

Operations Research and Information Systems are different scientific disciplines, yet they clearly have non-zero intersection. In this panel session, the four Editors- in-Chief of the UK OR Society journal “Health Systems” present their perspectives on the relationship between OR and IS, and how each can inform and enrich the other, especially in the domain of healthcare.

305