Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MB65

examples that by grouping similar ARMA demand streams together and forecasting several partial-aggregate streams the retailer is able to drastically improve its forecast accuracy when compared to grouping all consumer streams together and forecasting a single aggregate ARMA stream. Furthermore we show that little is lost when compared to forecasting all demand streams separately. 3 - Spatial Patterns and Socioeconomic Dimensions of Short-term Shared Accommodations: The Case of Airbnb in Los Angeles and New York City Avijit Sarkar, Professor, University of Redlands, 1200 E. Colton Avenue, Redlands, CA, 92373, United States, Mehrdad Koohikamali, James B. Pick This study examines spatiotemporal aspects and socioeconomic dimensions of shared accommodations within the broader context of the sharing economy. Specifically, we examine how socioeconomic attributes of Airbnb hosts moderated by hosts’ attitudes towards trust and greener consumption influence participation in the sharing economy. Spatial bias in sharing economy participation rates is examined and policy implications for the supply side of shared accommodations are discussed along with generalizability of results for two major U.S. cities - New York City, NY and Los Angeles CA. 4 - Cognitive Learning with Application to Supply Chains Seho Kee, Arizona State University, Tempe, AZ, 85044, United States, Ghazal Shams, Mani Janakiram, Mark A. Wilkinson, George Runger The evolving data-rich environment has enabled making data-driven business decisions. Rather than traditional approaches, more automated, adaptive, and intelligent systems can be developed to expand upon the capabilities of human decision makers. A cognitive computing approach is presented to learn and improve human decision-making in a dynamic, context-aware decision support platform. The proposed system is applied to a real-world use case in supply chain. 5 - Sourcing Model Optimization using a Linear Model in Multi-echelon Distribution Network Prashant Kaldindi Verma, FleetPride, Inc., 600 E. Las Colinas Blvd #400,, Irving, TX, 75039, United States, Mohit Arora, Homarjun Agrahari, Brian Steinmiller, Ziyu Li The SKU sourcing policy can have significant impact on the transportation and inventory holding costs. We developed and implemented an optimization model that explicitly deals with vendor contracts, varying pack sizes and demand. The model takes into account the cost of transportation to each location and handling cost specific to location. We will present mathematical model, and financial impact it has had on the organization. n MB65 West Bldg 104B Joint Session DM/Practice Curated: Data Science and Applied Probability Sponsored: Data Mining Sponsored Session Chair: Xu Sun, Columbia University, New York City, NY, 10027, United States 1 - Shepherding the Herd Ville Satopaa, INSEAD, 140 Avenue Daumesnil, Paris, 75012, France, Jussi Keppo We analyze multiple agents who forecast an underlying dynamic state based on a stream of public and private signals. Each agent minimizes a convex combination of her forecasting error and deviation from the other agents’ forecasts. As a result, the agents exhibit herding behavior - a bias that has been well-recognized in the literature. We first derive and analyze the agents’ optimal forecast under different levels of herding. This extends the Kalman filter to applications where herding is an important part of the process. After that we solve a dynamic strategy that allows a social planner to influence the agents’ forecasting and this way raise welfare through disclosure of public information. 2 - Determining Optimal Treatment Allocations for Online Experiments with a Fixed Number of Rounds Balaji PitchaiKannu, Research Scholar, IIT Madras, 236, Bhadra Hostel, IIT Madras, IIT Madras, CHENNAI, 600036, India, Nandan Sudarsanam In this study, we determine the optimal treatment allocation for experiments that are conducted in batches, in a live setting with a finite horizon. Unlike the traditional online learning problems seen in the multi-armed Bandit setup, our problem statement requires that the entire trial horizon is exhausted in a fixed set of rounds. In this environment, we determine the batch size as well as the ratio of treatments to units for each round. Our optimization seeks to minimize expected cumulative regret, a common metric in online learning. We use a Bayesian framework to model the theoretical means of the alternatives and the noise in the system as Gaussian distributions.

n MB63 West Bldg 103B

Joint Session DM/AI/Practice Curated: From Data to Algorithms: How Ride-sharing Platforms Improve Rider and Driver Experiences? Sponsored: Data Mining Sponsored Session Chair: Helin Zhu, UBER Co-Chair: Peter Frazier, Cornell University, School of Operations Research, and Information Engineering, Ithaca, NY 1 - Driver Preferences at Uber: Welfare, Flexibility, and Pricing Peter Frazier, Cornell University, School of Operations Research, and Information Engineering, Ithaca, NY, 14853, United States The Uber platform lets drivers set preferences for trips going to a destination through its Driver Destination feature. This feature lets drivers avoid deadheading when they need to go home, but also creates gaming opportunities by letting drivers select lucrative long trips. Moreover, when many drivers use this feature, their reduced flexibility causes riders going to unpopular destinations to wait longer on pickups for drivers coming from further away. We study the impact on welfare of this and other similar preference-setting features in two-sided markets. We show that they can decrease welfare when launched without discretion, but improve welfare when combined with appropriate pricing. 2 - Dynamic Pricing and Matching in Ride-hailing Platforms Helin Zhu, Marketplace Optimization, Uber, 1455 Market Street, 4th Floor, San Francisco, CA, 94103, United States, Chiwei Yan, Dawn Woodard, Nikita Korolko Advanced matching and dynamic pricing algorithms are the two key levers in ride-hailing platforms. We demonstrate via a stylized model, calibrated with real data, that the synergy of these two levers can bring significantly more benefits. Specifically, we study a novel matching mechanism called dynamic waiting. We show that pricing and waiting could be jointly optimized in reducing rider and driving waiting times, lowering price volatility, and increasing trip volume and welfare. 3 - Estimating Primary Demand in Bike-sharing Systems Chong Yang Goh, Massachusetts Institute of Technology-ORC, Cambridge, MA, 02139-4910, United States, Chiwei Yan, Patrick Jaillet We consider the problem of estimating the primary or first-choice demand for a bike-sharing service using trip and inventory data. To account for choice substitutions, we propose a rank-based demand model that treats each observed trip as the best available option in a latent ranking over origin-destination pairs. We then solve a high dimensional estimation problem using algorithms that (i) find sparse representations of the parameters efficiently, and (ii) constrain trip substitutions spatially according to the bike-share network. Our method is effective in recovering the primary demand and computationally tractable on a city scale, as we demonstrate on a bike-sharing service in Boston.

n MB64 West Bldg 104A Joint Session DM/Practice Curated: Data Science and Analytics in Supply Chain Sponsored: Data Mining Sponsored Session Chair: Aihong Wen, Walmart 1 - Reinforce Learning with Counterpetition Strategy in Hotelling Model

Xiaoya Xu, Assistant Professor, Guangdong University Of Finance & Economics, 21 Luntou Road, Guangzhou 510320, Guangdong, P.R. China, Guangzhou, 510320, China The combination of reinforce learning and game theory is exceptional serviceable for solving complexity problems, we adapt this methodology to the cooperation game in hotelling model. Assuming customers apply their potential game in each stage for maximizing their optimal revenue. Using the reinforce learning, customers can automatically choosing their optimal strategy systematically . 2 - Improving Demand Forecast Accuracy through Partial Aggregation Vladimir Kovtun, Assistant Professor, Yeshiva University, 500 W. 186th Street, New York, NY, 10033, United States, Avi H. Giloni, Sridhar Seshadri, Hurvich Clifford We consider a single firm (retailer) that observes demand streams coming from many distinct consumer populations. We assume that each demand stream follows an Autoregressive Moving Average (ARMA) process. We show with

171