Informs Annual Meeting 2017

WB03C

INFORMS Houston – 2017

2 - Off-policy Evaluation and Optimization with Continuous Treatments Angela Zhou, Cornell University, Ithaca, NY, United States, az434@cornell.edu, Nathan Kallus We develop new methods for off-policy evaluation and learning with continuous treatments, going beyond previous work on discrete treatments. Using observational data such as electronic health and transactional records, our method is able to learn effective policies in such contexts as medical dosing and personalized discounts. Our estimator uses a kernel to smoothly reweight data by the proximity of proposed treatments and performs well in a case study on personalized warfarin dosing. We analyze the bias and error of our method and prove favorable regret bounds that support this empirical success. 3 - Efficient Policy Learning Stefan Wager, Stanford GSB, Stanford, CA, 94305, United States, swager@stanford.edu There is considerable interest in methods that reduce the problem of learning good treatment assignment policies to the problem of accurate policy evaluation. Given a class of candidate policies, these methods first effectively evaluate each policy individually, and then learn a policy by optimizing the estimated value function. Despite the wealth of proposed methods, the literature remains largely silent on questions of statistical optimality, and there are only limited results characterizing which policy evaluation strategies lead to better learned policies than others. Here, we consider efficiency bounds for policy learning, and propose new methods motivated by this optimality theory. 4 - Sensitivity Analysis for the Sample Average Treatment Effect under Effect Heterogeneity In observational studies, a sensitivity analysis assesses how strong unmeasured confounding needs to be to alter the study’s findings. Methods for sensitivity analysis have been established under the assumption that the treatment effect is constant for all individuals. While innocuous for the analysis of randomized experiments, it has been argued in observational studies that certain patterns of unobserved effect heterogeneity may render the performed sensitivity analysis inadequate. We present a method for conducting a sensitivity analysis in the presence of effect heterogeneity. In so doing, we illustrate that concerns about the constant treatment effect model are largely unwarranted. WB03B Grand Ballroom B Empirical/Data Driven Revenue Management and Pricing Sponsored: Revenue Management & Pricing Sponsored Session Chair: Serguei Netessine, University of Pennsylvania, Wharton School, Philadelphia, PA, 138676, United States, netessin@wharton.upenn.edu Co-Chair: Jun Li, Ross School of Business, University of Michigan, Ann Arbor, MI, 48109, United States, junwli@umich.edu 1 - How to Operationalize Modern Personalization Algorithms Vivek Farias, MIT, 100 Main Street, Cambridge, MA, 02142, United States, vivekf@mit.edu, Andrew A.Li, Deeksha Sinha Modern recommender systems rely on encoding preferences via a mapping of users and items to a latent space, and while there exist sophisticated methods to estimate such mappings, the problem of efficiently operationalizing these embeddings is not well-understood. Thus motivated, we propose a sublinear time algorithm for this problem, and show that it achieves a constant-factor approximation under general random utility choice models. Our algorithm relies on a new sublinear time sampling scheme, which we develop to solve a class of problems that subsumes the classic nearest neighbor problem. We demonstrate the efficacy of our approach on a massive content discovery dataset from Outbrain. 2 - A Conditional Gradient Approach for Nonparametric Estimation of Mixtures of Choice Models Ashwin Venkataraman, New York University, 719 Broadway, 7th Floor, New York, NY, 10003, United States, ashwin@cs.nyu.edu, Srikanth Jagabathula, Lakshminarayanan Subramanian We propose a new method to nonparametrically estimate the mixture of logit models from aggregated sales transactions and product availability data. We formulate the estimation problem as an infinite-dimensional constrained convex program and use the conditional gradient (aka Frank-Wolfe) algorithm to solve it. We characterize the structure of the estimated mixing distribution. In two real- world case studies, we outperform the EM benchmark on speed (16x faster), in-sample fit (up to 24% gains in the log-likelihood loss), and predictive (28% gains on the RMSE metric in predicting market shares) and decision accuracies (extracts 1.4x more revenue). Colin Fogarty, Massachusetts Institute of Technology, Cambridge, MA, 02114, United States, cfogarty@mit.edu

3 - Discrimination with Incomplete Information in the Sharing Economy: Field Evidence from Airbnb Dennis Zhang, 8342 Delcrest Drive, Apt 328, University City, MO, 63124, United States, denniszhang@wustl.edu, Ruomeng Cui, Jun Li Recent research has found widespread discrimination by hosts against guests of certain races in online marketplaces, which endangers the very basis of a sharing economy — building trust in the communities. In this paper, we explore the root cause of discrimination and how to eliminate discrimination by conducting two

randomized field experiments among 1,256 hosts on Airbnb. 4 - Multi-dimensional Decision Making in Operations: An Experimental Investigation of Joint Pricing and Quantity Decisions

Necati Tereyagoglu, Assistant Professor, Georgia Institute of Technology, 800 W Peachtree St NW, Atlanta, GA, 30308, United States, necati.tereyagoglu@scheller.gatech.edu, Karthik Ramachandran, Yusen Xia Firms in several industries like medicine, retail and publishing must jointly determine the price and production quantity of their products well in advance of the selling season. In this study, we experimentally examine subjects’ performance when they jointly determine price and quantities. We find that subjects systematically deviate from the theoretically optimal price and quantity levels. In a series of follow-up experiments, we isolate several factors that influence decision-making in this context. We show that decisions improve by making subjects more aware of interdependence and by reducing the uncertainty.

WB03C Grand Ballroom C

Customer Choice in Retail Operations Sponsored: Manufacturing & Service Oper Mgmt Sponsored Session

Chair: Adam J. Mersereau, University of North Carolina, Chapel Hill, NC, 27599-3490, United States, ajm@unc.edu 1 - Online Assortment Optimization when Customers Refine Their Search Zhichao Feng, University of Texas at Dallas, 2200 Waterview Pkwy, Apt 1828, Richardson, TX, 75080, United States, zxf140830@utdallas.edu, Shengqi Ye, Dorothee Honhon When shopping online, a consumer often searches a keyword and checks the products displayed by the retailer. In many cases, the retailer has numerous products matching the keyword, with different features and functionalities, but is only able to show a subset of them due to limited displaying space. The assortment shown by the retailer influences the consumer’s decision to buy or not. In addition, when the consumer is not familiar with the product category, the assortment may trigger interest in a specific product feature, leading the consumer to refine her search, and focus only on products with this feature. Taking this into consideration, we study the online retailer’s optimal assortment decision. 2 - On Consumption and Composition of N-packs John H. Semple, SMU, 6212 Bishop Blvd, Dallas, TX, 75205, United States, jsemple@mail.cox.smu.edu, Aydin Alptekinoglu We consider the optimal consumption of a pack of n substitutable products within a general random utility framework. Using dynamic programming, we demonstrate that many basic properties of this problem are independent of the error distribution. 3 - Learning Choice Model Trees: Recursively Segmenting Individuals to Better Understand Their Choices Michael Young, Doctoral Candidate, Harvard Business School, Choice modeling is a technique used by many retailers to better understand the impact of operational levers, e.g. price, on the customer’s product choice. Similarly, customer segmentation is a common practice used to personalize the customer’s shopping experience, e.g. targeted ads/offers. In this work, we combine a popular machine learning technique with a choice model in order to jointly achieve both choice modeling and customer segmentation objectives. Boston, MA, United States, miyoung@hbs.edu, Kris Johnson Ferreira, Mohammed Ali Aouad

473