Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SB20

2 - A Choice Modeling Framework for Service Time Windows Xiao Lei, Columbia University, 500 W. 120th St., New York, NY, 10027, United States, na, Adam Elmachtoub On-demand services have become increasingly common, and typically allow customers to choose a time window to receive the service. As a result, there is a natural trade-off between on-time customer service and operational cost. To address this issue, some service providers offer large time windows with rewards, together with the normal small ones. In this paper, we provide a choice modeling framework to address how customers choose among time windows, and apply this framework to evaluate various strategies for time window design. 3 - When Behavioral Operations Meets Optimization: Evidence From a Large Field Experiment on Behavioral Bin-packing Algorithms Jiankun Sun, Northwestern University, Evanston, IL, United States, Dennis Zhang, Haoyuan Hu, Jan Van Mieghem In logistics, optimization algorithms are widely deployed to augment and empower human. However, human may not execute the algorithm solutions for various reasons, which might cause efficiency loss and cost increase. We examine how a behavioral optimization algorithm affects human’s compliance rate of algorithm solutions, working efficiency and estimate the corresponding economic value by running a field experiment on two bin-packing algorithms. The first algorithm is a conventional heuristic bin-packing algorithm, while the second algorithm extends the first one by considering human behavior in choosing the packing boxes. 4 - Revenue Management versus Machine Learning: Finding Optimal Product Displays at Alibaba Jacob Feldman, Olin Business School, 6 Portland Court, Saint Louis, MO, 63108-1291, United States, Dennis Zhang We compare the performance of two state of the art approaches for finding the optimal set of products to display to customers arriving to Alibaba. The first procedure embeds hundreds of product and customer features within a sophisticated machine learning algorithm that is used to estimate the purchase probabilities of each product for the customer at hand. The products with largest expected revenue (revenue*predicted purchase probability) are then made available for purchase. Our second approach uses a featurized MNL model to predict purchase probabilities for each arriving customer and then solves cardinality constrained assortment optimization problems. Data Science in Online Platforms Sponsored: Revenue Management & Pricing Sponsored Session Chair: Ruben Lobel, Airbnb, San Francisco, CA, 94103, United States Co-Chair: Carlos Abad, Airbnb, San Francisco, CA, 94103, United States 1 - Supply-side Incentives Allocation at Lyft Davide Crapis, Lyft, San Francisco, CA, United States Lyft runs a number of programs to engage drivers when and where supply is most needed. The aim is to align supply and demand, provide better service levels for passengers and increased earnings for drivers. Scientists work on optimizing these programs using machine learning, optimization algorithms, and continuous experimentation. We give an overview of different incentives programs and the technical problems they pose. We then present in more detail some of the techniques used to compute optimal allocations. 2 - Zillow Instant Offers with Human-in-the-loop Machine-learning David E. Fagnan, Senior Manager, Applied Science, Zillow Group, Seattle, WA, United States, Fan Cao, Sebastian Wickenburg, Emily Gill, Krishna Rao We give an overview of recent data science challenges at Zillow. In particular, we focus on Zillow Instant Offers, a new home-selling option for homeowners who want a certain and predictable sale on their timeline. Our system tries to balance offer quality, offer timeliness, and risk in the presence of potential selection bias. With this focus, we discuss our learnings around making decisions under uncertainty including examples such as how much to pay for a house, predicting housing liquidity through days on market, and how much a pool is worth. We discuss progress on these problems and how we combine the use of human judgement with machine-learning algorithms in a human-in-the-loop system. n SB22 North Bldg 130 Joint Session RMP/Practice Curated:

n SB20 North Bldg 129A Revenue Management for Marketplaces Sponsored: Revenue Management & Pricing Sponsored Session Chair: Omar Besbes, Columbia University, New York, NY, 10027, United States Co-Chair: Ilan Lobel, New York University, New York University, New York, NY, 10012, United States 1 - Frustration-based Promotions: Field Experiments in Ride Sharing Maxime Cohen, NYU Stern, Baek Jung Kim, Michael-David Fiszer In this talk, we examine whether a firm should proactively send compensation to users who have experienced a frustration (i.e., a poor service quality). In collaboration with one of the leading ride-sharing platforms, Via, we designed and ran three field experiments to investigate how different compensation types affect the engagement of riders who experienced a frustration. 2 - Position Ranking and Auctions for Online Marketplaces Heng Zhang, USC Marshall School of Business, Bridge Memorial Hall - BRI 401B, 3670 Trousdale Parkway, Los Angeles, CA, 90089, United States, Leon Yang Chu, Hamid Nazerzadeh We study how online e-commerce platforms should rank products displayed to consumers, and utilize the top slots. We present a model that considers consumers’ search costs and the externalities sellers impose on each other, which allows us to study a multi-objective optimization, whose objective includes consumer, seller surplus, and the sales revenue, and derive the optimal ranking decision. In addition, we propose a surplus-ordered ranking mechanism, motivated in part by Amazon’s sponsored search program, for selling top slots. We show that our mechanism is near-optimal, performing significantly better than those that do not incentivize the sellers to reveal their private information. 3 - Surge Pricing and Its Spatial Supply Response Francisco Javier Castro, Columbia University, Columbia School of Business, 527 West 121st, New York, NY, 10027, United States, Omar Besbes, Ilan Lobel We study the pricing problem faced by a platform matching price sensitive customers to flexible supply units in a city. The platform sets prices, and drivers react by choosing where to travel based on prices, travel costs and congestion levels. By uncovering an appropriate knapsack structure to the platform’s problem, we first establish a characterization of the optimal solution and its supply response. We then tailor the analysis to a demand shock and derive in quasi-closed form the optimal solution. The platform uses prices to create damaged regions where demand is shut-down or driver congestion is artificially high, incentivizing some drivers to travel both toward the demand shock and away from it. 4 - Value Loss in Allocation Systems with Provider Guarantees Xavier Warnes, Stanford University, Yonatan Gur, Dan Andrei Iancu Centralized planning systems that allocate tasks to workers or service providers must often restrict their allocations so as to ensure particular (welfare) guarantees to their workers. Such restrictions can generate losses in the total value created or in the system’s share of that value. Our work provides a uniform bound for these losses under a very broad class of restrictions due to worker guarantees. The bound only depends on the number and the heterogeneity of the service providers, and allows identifying the guarantees that are most stringent. We also show that such value losses remain small in practical settings of interest calibrated with real data.

n SB21 North Bldg 129B

Data Driven Revenue Management Sponsored: Revenue Management & Pricing Sponsored Session

Chair: Jake Feldman, Washington University in St. Louis 1 - Pricing Discounts in Electric Vehicle Share Systems Bobby Nyotta, Fernanda Bravo, Jacob Feldman

We study the use of pricing discounts in an electric vehicle share system with free-floating parking. Specifically, we examine the optimal policy for the operator to offer free rides to the charging stations. Free rides can help to keep vehicles charged and to rebalance the system, however a revenue loss is experienced. Based on the optimal policy structure, we decompose and approximate the problem with an algorithm that performs well on real problem instances.

38