Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MA52

4 - A Parallel Machine Scheduling Problem with Release Dates, Equal Processing Times, and Eligibility Constraints Kangbok Lee, Associate Professor, POSTECH, 77 Cheongam-Ro. Nam-Gu., POSTECH, Industrial & Management Engineering, Pohang, 37673, Korea, Republic of, Juntaek Hong We consider a problem of scheduling independent jobs on parallel machines to minimize the total completion time. Each job has a given release date and a set of eligible machines, and all jobs have equal processing times. For the problem with a fixed number of machines, we show its computational complexity of the problem by providing a polynomial time dynamic programming algorithm. For the problem with an arbitrary number of machines, we propose approximation algorithms and they are evaluated by worst-case analysis and numerical experiments. n MA52 North Bldg 231C Social Media and Online Communities Emerging Topic: Social Media Analytics Emerging Topic Session Chair: Sung Won Kim, University of Illinois at Urbana Champaign, Champaign, IL, 61820, United States 1 - Market Sensing from User Review for Product Management Ujjal Kumar Mukherjee, Assistant Professor, University of Illinois, Urbana-Champaign, 306 Wohlers Hall, 1206 South Sixth Street, Champaign, IL, 61820, United States In this paper, we use social media data on customer reviews to understand how firms can use this data to sense market reaction to their products and services. We use sentiment analysis and text analysis on customer feedback to show that customer sentiment significantly predicts future market share of products. We also compare and contrast several methods of customer feedback analysis. 2 - Similarity in the Crowd(funding): How Funder and Backer Affinity Contributes to Crowdfunding Choices Lauren Rhue, Wake Forest University, 1834 Wake Forest Rd, 212 Farrell Hall, Building 60, Winston-Salem, NC, 27106, United States, Jessica M. Clark, Lauren Dahlin Using Kickstarter data, we capture funder and backer attributes to examine whether they share similar characteristics and whether those similarities shape the propensity to contribute to particular projects. We also examine project similarity to uncover whether backers give to projects in particular categories or communities. 3 - Unsupervised Dimensionality Reduction vs. Supervised Regularization for Classification from Sparse Data Jessica Clark, University of Maryland, College Park, MD, United States, Foster Provost There is consensus among existing guidelines that supervised regularization is superior to unsupervised Dimensionality Reduction (DR) techniques for mitigating overfitting in predictive modeling; however, these guidelines do not take into account that the two types of techniques are often used in conjunction. Many published studies using DR for applied data mining fail to compare performance using the original feature set. We experimentally compare binary classification performance using DR features versus original features under numerous conditions, and find that generally, DR does not add value beyond supervised regularization, and can even diminish performance. n MA53 North Bldg 232A Joint Session AMD/RMP: Algorithms and Market Design Sponsored: Auction and Marketing Design Sponsored Session Chair: Robert Day, University of Connecticut, Storrs, CT, 06269-1041, United States 1 - Learnability and Models of Decision Making under Uncertainty

2 - Learning and Efficiency in Games with Dynamic Population Thodoris Lykouris, PhD Candidate, Cornell University, 107 Hoy Road, Gates 336, Ithaca, NY, 14853, United States, Vasilis Syrgkanis, Eva Tardos We study multi-player game settings (such as online advertising, internet routing, and bandwidth allocation) where the player set is dynamically evolving over time and participants use some learning algorithms to adapt to the changing environment. Traditional equilibrium notions require too much information for the players since they need to form perfect beliefs about the behaviors of others and do not extend to dynamic settings. In contrast, we show that, under the learning behavioral assumption, players reach outcomes of high quality (measured by the social welfare) even when there is a high churn in the population. 3 - Selling to a No-regret Buyer Jieming Mao, Princeton University, Princeton, NJ, United States, Mark Braverman, Jon Schneider, Matt Weinberg We consider the problem of a single seller repeatedly selling a single item to a single buyer (specifically, the buyer has a value drawn fresh from known distribution D in every round). Prior work assumes that the buyer is fully rational and will perfectly reason about how their bids today affect the seller’s decisions tomorrow. In this work we initiate a different direction: the buyer simply runs a no-regret learning algorithm over possible bids. We provide a fairly complete Juba Ziani, California Institute of Technology, Pasadena, CA, United States, Yiling Chen, Nicole Immorlica, Brendan Lucier, Vasilis Syrgkanis We consider a data analyst’s problem of purchasing data from strategic agents to compute an unbiased estimate of a statistic of interest. Agents incur private costs to reveal their data and the costs can be arbitrarily correlated with their data. We design an individually rational and incentive compatible mechanism that optimizes the worst-case (over the unknown correlation between costs and data) mean-squared error of the estimation, subject to a budget constraint. We give the optimal mechanism in closed-form. We extend our results to acquiring data for estimating a parameter in regression analysis, where private costs can only correlate with the values of the dependent variable. n MA54 North Bldg 232B Leveraging Human Intelligence Alongside Analytical Models Sponsored: Behavioral Operations Management Sponsored Session Chair: Blair Flicker, University of Texas-Dallas, Richardson, TX 1 - Managerial Insight and “Optimal” Algorithms Blair Flicker, University of Texas-Dallas, Richardson, TX, 75208, United States Stylized models omit many real-world phenomena, and such model misspecification degrades the performance of “optimal” policies in practice. By accounting for unmodeled dynamics, human managers can improve decision making. I formalize “managerial insight” (noisy signal of demand observable only by humans) and apply it to the newsvendor setting. With superior demand information, human newsvendors should improve profits, but experiments reveal costly suboptimal ordering. I recast the newsvendor task as a forecasting task and develop an algorithm to convert forecasts to orders. In the lab, this human- machine hybrid approach consistently outperforms either humans or machines operating alone. 2 - Can We Improve Analytical Models through Judgement and Local Information Han Oh, Texas A&M University, College Station, TX, 77843, United States, Rogelio Oliva Using data from a large retailer, we identify the triggers for managers modifying re-stocking order recommendations from a centralized information system and evaluate the performance of such modifications. Our analysis shows that local managers do have local information that adds value to the overall decision making - thereby improving the performance of a centralized information system. 3 - When Models Meet Managers: Integrating Judgmental and Statistical Model-Based Forecasting John Aloysius, University of Arkansas, Supply Chain Management Department, Wcob 204, Fayetteville, AR, 72701, United States, Enno Siemsen, Rebekah Brau Technology-enabled big data analytics and machine learning is increasingly a feature of forecasting in practice. However, practitioners continue to use judgment to incorporate private information to improve the accuracy of forecasts generated by statistical models. Our research examines how statistical models and human judgment may be integrated to improve forecast accuracy in different forecasting environments. characterization of optimal auctions for the seller in this domain. 4 - Optimal Data Acquisition for Statistical Estimation

Pathikrit Basu, California Institute of Technology, 1633, Amberwood Drive, South Pasadena, CA, United States, Federico Echenique

We study whether some of the most important models of decision-making under uncertainty are uniformly learnable, in the sense of PAC (probably approximately correct) learnability. Many studies in economics rely on Savage’s model of (subjective) expected utility. The expected utility model is known to predict behavior that runs counter to how many agents actually make decisions (the contradiction usually takes the form of agents’ choices in the Ellsberg paradox).

138