Symposium on Data-Driven Decision Making: Modeling, Algorithms and Implementations

MAY 2, 2019

SESSION 1: 8:45 AM - 10:15 AM

Srikanth Jagabathula is a Visiting Associate Professor in the Technology and Operations Management (TOM) Unit at the Harvard Business School. He is also Associate Professor (on leave) of Information, Operations, and Management Sciences at New York University’s Leonard N. Stern School of Business.

His research interests broadly lie at the intersection of operations, machine learning, and marketing. The objective of his research is to obtain easy-to-use techniques for a wide range of managerial decisions: the right products to design, the right products and prices to offer to customers, and the right quantity of each product to carry. His work on assortment planning and customer segmentation has been implemented at various firms, such as The Bon-Ton, Ford Motor Company, and, leading to demonstrable improvements. He has also worked with, Microsoft Research, and Deutsche Bank. He has received a number of awards recognizing his work, including the NSF CAREER Award, the Wickham Skinner Early-Career Research Accomplishments Award from the Production and Operations Management Society, best student paper awards in operations and machine learning conferences, best master’s thesis award, the IIT Bombay President of India Gold Medal in 2006, and the 2018 Poets & Quants “40 Best Business Professors Under 40.” Professor Jagabathula received a B.Tech. degree in Electrical Engineering from IIT Bombay, and an S.M. degree and Ph.D. in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology.

•Abstract: Personalizing Retail Promotions through a DAG-based representation of customer preferences
Price and display promotions are common in retail settings. For instance, an estimated 45% of Safeway’s sales in 2013 came through specialized offers. With the proliferation of technology options, retailers (both online and brick-and-mortar) are now moving from mass promotions to personalizing promotions to individual customers. For instance, in the offline brick-and-mortar stores, retailers can launch personalized coupons on mobile phones.

When personalizing promotions, the retailer should not promote an item that the customer will buy anyway, but at the same time use the opportunity to switch customer’s purchase to other products. Carrying this out optimally requires a nuanced understanding of customer preferences over different brands.

To solve this problem, we propose a new representation of customer preferences in the form of a directed graph. A directed edge A -> B from brand A to brand B in the graph indicates that the customer strongly prefers brand A over brand B, and this preference remains stable over time. We assume that these preferences are acyclic, so our representation will become a directed acyclic graph (DAG). We propose techniques using which a retailer can infer the DAG representation of each customer from observed purchase transactions.

On data from a collection of grocery retailers, we show that the DAG representation provides a powerful way for the retailer to reason through promotion decisions. For instance, if the brand A is already the most preferred of all the available brands on the shelf, then the retailer shouldn’t promote brand A. If the retailer wants to induce purchase of brand C, then the retailer should not only promote brand C but also not promote brand B. We propose an algorithm that does this reasoning automatically. Using this algorithm, we find that the retailer can increase its profits by more than 11% compared to its existing promotion policy.

Tunay Tunca, University of Maryland
Tunay Tunca is a Professor of Management Science and Operations Management at Robert H. Smith School of Business at University of Maryland. Prior to joining University of Maryland, he was an Associate Professor of Operations, Information, and Technology at Stanford GSB, where he received his Ph.D. from. Professor Tunca has also been a visiting scholar at the Sloan School of Business at Massachusetts Institute of Technology, Wharton School of Business at University of Pennsylvania, Yahoo Inc., and Hewlett-Packard. His research interests are in economics of operations management, focusing on theoretical and empirical analysis of supply chains, innovative operations, and sharing economy business models. His work has won awards and recognitions from Management Science, M&SOM, POMS, and CSAMSE. He is the winner of several teaching awards at the Robert H. Smith School of Business, including the 2014 Allen J. Krowe Teaching Excellence Award, and Distinguished Teaching Awards in 2015-2018. He currently serves as an Associate Editor for the journals Management Science and M&SOM.

•Abstract: An Empirical Analysis of Market Formation, Pricing, and Revenue Sharing in Ride-Hailing Services
We study the two-sided market for ride-hailing services and the effects of pricing policies and regulation on the value generated by these platforms. We first develop a discrete choice model for the formation of mutually dependent demand (consumer side) and supply (driver side) that jointly and endogenously determine pricing. Using this model and a comprehensive data set obtained from the largest mobile ride platform in China, we estimate consumer and driver price elasticities as well as other factors that affect market participation for the two main markets the company hosts, namely basic ride-hailing and taxi services, with a high level of model fit. Based on these estimation results and counterfactual analysis, we demonstrate that surge pricing improves consumer and driver welfare as well as platform revenues, while reducing taxi revenues on the platform. However, our analysis suggests that surge pricing can hurt consumer surplus and platform revenues during non-peak hours, and hence should be used selectively during the day. We show that platform revenues can be maximized by increasing drivers' revenue share from the current levels. We measure the effects of proposed regulations such as imposing ride-hail price hikes, and demonstrate that they would decrease consumer and driver surplus significantly. Finally, we estimate that the platform's basic ride-hailing services generated consumer value equivalent to more than 32 Billion USD in China in 2017.

SESSION 2: 10:30 AM - NOON

Qi (Annabelle) Feng is the John and Donna Krenicki Chair in Operations Management at Krannert School of Management, Purdue University. She was previously a faculty member at McCombs School of Business, The University of Texas at Austin. She received her Ph.D. in Operations Management from UT Dallas in 2006. Her main research interest lies in studying firms’ sourcing decisions in the broad context of supply chain management. Her work focuses on individual firm’s procurement planning in uncertain environment and multiple firms’ interactions in sourcing relationships. She also works in the areas of product development and proliferation management, resource planning, economic growth models, and information system management. She is currently a Department Editor of Data Science, Stochastics and Optimization for Production and Operations Management, the Department Editor of Supply Chain Management for Flexible Manufacturing and Service Journal, and an Associate Editor for Management Science. She received the first prize in the INFORMS Junior Faculty Paper Competition in 2009, the Franz Edelman Award in 2009, the Wickham Skinner Early-Career Research Accomplishment Award in 2012, and the Wickham Skinner Best Paper Award in 2018.

•Abstract: Operational Data Analytics: A Newsvendor Application
With the development of computing technology and data availability, an increasing attention is paid to data-driven and data-integrated decision making in practice and research. We propose a framework of Operational Data Analytics (ODA) and demonstrate its application through the example of newsvendor model. In this talk, we focus on the situation where we have full structural knowledge but may be uncertain about the statistical characterization of the model. The ODA framework integrates data into decision making by carefully formulating the data-integration model and the validation model. The data-integration model identifies an appropriate class of operational statistics (i.e., the decision as a statistics of the data) and the validation model finds the best performer among the operational statistics. In contrast to the traditional estimation-optimization approach, which does not optimize the actual performance measure, the ODA approach significantly improves decision quality especially when the data is limited.


Stefanus Jasin is an Associate Professor of Technology and Operations at the Ross School of Business, University of Michigan, Ann Arbor. He earns his PhD in Computational and Mathematical Engineering from Stanford University and is broadly interested in topics that lie at the intersection of OR, OM, IS, and Marketing, with an emphasize on developing provably near-optimal and easily implementable heuristic policies for complex (and often large-scale) dynamic control problems. Some of his works include: real-time pricing, e-commerce order fulfillment, delivery consolidation, inventory optimization, and joint learning and optimization. Most recently, he is also working on optimization in on-demand market.

•Abstract: Randomized Product Framing, Pricing, and Order Fulfillment for E-Commerce Retailers
How should an online retailer decide the set of products to be displayed for promotion, the set of products to be offered with a free shipping or a free one day delivery, or the set of products to be placed at the top of a display page? These are examples of so-called "product framing". It has been widely noted in the empirical literature that product framing matters. They affect customers' attention, which in turn affect their purchasing decision. This talk presents a recent joint work on the topic of randomized product framing, pricing, and order fulfillment for e-commerce retailers. There is a finite selling season and each product has a limited inventory, with no replenishment opportunity.

We develop a heuristic policy based on the solution of a deterministic approximation of the original stochastic control problem and show that it is asymptotically optimal. The key challenge here is that the solution of the deterministic problem is not directly implementable and needs to be carefully transformed into a "feasible solution", which we eventually interpret as a randomized policy. We numerically test our heuristic policy using both synthetic and real-world data provided by a major US retailer. The results show that the proposed heuristic is very close to optimal and also outperforms some state-of-the-art algorithms on other dimensions.

SESSION 3: 1:00 PM - 2:30 PM

Dr. Morris A. Cohen is the Panasonic Professor of Manufacturing and Logistics in the Operations, Information and Decisions Department of The Wharton School of the University of Pennsylvania. He is also Co-Director of Wharton’s Fishman-Davidson Center for Service and Operations Management. Morris has been a visiting faculty member at the Indian Institute of Management in Bangalore, the Shanghai Institute of Foreign Trade, the Technion in Israel, Kobe University in Japan, the Interdisciplinary University in Israel, Stanford University, Shanghai Jiao Tong University, and the Massachusetts Institute of Technology. He also has served as an editor for a variety of journals; including Operations Research, Journal of Operations Management, Journal of Manufacturing and Operations Management, Naval Research Logistics, Supply Chain Management Review, the Journal of Production and Operations Management, and Service Science. He is currently the department editor for services in the Journal of Manufacturing and Service Operations Management. He is a Fellow of the Institute for Operations Research and Management Science, and a Senior Fellow of the Manufacturing and Service Operations Management Society. He is a recipient of the 2015 Wharton School Teaching Commitment and Curricular Innovation Award, the 2007 Tibbets Award from the National Science Foundation and the 2001 Institute of Industrial Engineers Award for the best paper published in IIE Transactions in 1999. He also has been a plenary speaker at the Japanese Operations Management and Strategy Association (2013), the German Operations Research Society (2008) and the Production and Operations Management Society (2007).

Morris’s research interests include analysis of the current drivers of global supply chain sourcing strategy and product-service system modeling with a focus on performance-based incentives and buyer-supplier coordination to support a Servicization strategy. He has also developed and implemented advanced optimization tools for supply chain resource planning. His recent application work includes development of strategic and tactical planning systems for service supply chains in industries such as aerospace & defense, consumer electronics, health care technology, oil and gas, automobile, semiconductor equipment, computers, and telecommunications. He also has been a consultant to more than 50 multinational companies in the computer, semiconductor, automobile, food, military logistics, process equipment, pharmaceutical, capital equipment leasing, and industrial paper/plastics industries.

Until recently Morris was founder and chair of the board of MCA Solutions, a software company specializing in after-sales logistics planning systems, which recently merged with PTC, a leading provider of product design and service life cycle management decision support systems. He also has been a member of the Board of Directors of Vlasic Foods International, Inc., the Engineering Advisory Committee of the National Science Foundation and the Advisory Board for the Center for Transportation & Logistics, Massachusetts Institute of Technology. He recently founded a startup (D3 Analytics), that is applying concepts of machine learning and big data to a new paradigm for supply chain planning and control. He has also been a policy analyst for the planning branch of the Treasury Board Secretariat of the Government of Canada. Morris holds a B.A.Sc. in Engineering Sciences from the University of Toronto, as well as an M.S. in Industrial Engineering and a Ph.D. in Operations Research from Northwestern University.

•Abstract: Optimization for the Delivery of Service Differentiation Through Operating Segments: Design and Control Challenges
This talk presents a model hierarchy for developing an optimal differentiation strategy for the delivery of services. It is based on the framework introduced in Guajardo and Cohen [2018] for the management of value-added services that are bundled with manufactured products, (aka servicization). The model utilizes the concept of operating segments (introduced by Frei and Morriss [2012]) and considers the definition of market segments appropriate for differentiated services and the design of such services as well as the infrastructure required to produce them. The model integrates these design decisions with policies for the management of the operational processes required to deliver differentiated service products to multiple market segments. The model formulation incorporates tradeoffs, risks and constraints associated with decisions at all levels in the planning and control hierarchy and considers their interactions. Application of the model framework to the case of after-sales services is examined to illustrate how each stage of the methodology can be implemented. We also discuss model formulations that determine optimal service targets and the use of such targets to determine optimal resource rationing. The talk concludes with consideration of implementation challenges.


Scott Webster is the Bob Herberger Arizona Heritage Chair in Supply Chain Management at the W.P. Carey School of Business, Arizona State University. He serves as a senior editor for Production and Operations Management. His research focuses primarily on problems related to pricing and inventory, and on challenges that arise in agriculture and nonprofit settings.

•Abstract: Agricultural cooperative pricing of premium product)
We consider the problem of price setting by a cooperative for purchase of an agricultural product with the following characteristics: (1) the pricing policy is published prior to the growing season and (2) the value of the farmer’s harvest is affected by his investments during the growing season. To be viable, the price policy must be competitive with the open market from both the perspective of the farmer and the cooperative. One alternative is to set prices that mimic the open market prices, an approach that we find in use at major cooperatives we studied. There is sparse literature on inefficiencies in agricultural supply chains that arise from payment terms. New approaches that help reduce inefficiencies can affect the world’s food supply and living standards. Our work is motivated by evidence of farmer underinvestment in crop quality in the olive oil industry. We define and analyze a model of this system. We find that farmers consistently underinvest in crop quality under the open-market pricing policy. There are two drivers of this effect: (1) the degree to which the cooperative can command a higher retail price than a farmer, (2) farmer risk aversion. We propose two alternative pricing policies that are new to the agricultural literature, both of which can coordinate farmer decisions with the system but differ in terms of ease of implementation and susceptibility to risk aversion. We identify an easy-to-implement policy that can lead to meaningful gain when introduced in conjunction with crop insurance. We calibrate our model using data from the olive oil industry in Turkey and find a profit improvement of 10-15% over the current open-market pricing approach.

SESSION 4: 2:45 PM - 4:15 PM

Van-Anh Truong joined the Industrial Engineering and Operations Research Department in 2010. She received a Bachelor's degree from University of Waterloo in Mathematics in 2002, and a Ph.D. from Cornell University in Operations Research in 2007. Before coming to Columbia, she was a quantitative associate at Credit Suisse, and a quantitative researcher at Google.

She is interested in a broad class of problems that arise in Supply Chain Management, Healthcare, and Business Analytics. These problems address decision making under uncertainty in information-rich and highly dynamic environments. Her recent work focuses on real-time optimization approaches for large e-commerce, healthcare, and service applications. Her research is supported by an NSF Faculty Early Career Development (CAREER) Award. She is currently an Associate Editor for Operations Research and IIE Transactions.

•Abstract: Dynamic Optimization of Mobile Push Advertising Campaigns
We study a novel resource-allocation problem faced by Alibaba Group. In this problem, mobile “push messages” must be sent over the course of a day to hundreds of millions of users. Each message can be sent to any number of users, and yields a reward when it generates a clickthrough, subject to a budget constraint on the total reward over all users for the message. This budget represents the maximum amount that an advertiser is willing to pay for clickthroughs for the message on a given day. Given users’ diverse preferences, the problem aims to deliver the “right messages” to the “right users” to maximize ad revenues without overwhelming each user with too many messages. Due to the large size of the real application, we analyze algorithms for the above problem in an asymptotic regime. We consider a novel scaling of the problem “size,”called big-data scaling.

In this scaling, as the problem size grows, the number of users, as well as their diversity, grow. The scaling captures the fact that individual user information remains highly granular and distinctive even as the size of the user base increases. We prove that solving the problem as a static assignment problem results in a regret of O( √t), where t is the parameter scaling the problem. Furthermore, adding a single recourse opportunity, by sending push messages in two cycles over the course of a day and making use of information observed in the first cycle to adapt decisions in the second cycle, can reduce the regret to O(t 1/4 log t). Finally, the difference in regret between the static and dynamic strategy can be Ω(√ t). Numerical experiments on three real data sets, each containing several hundred million users, show that the latter strategy improves the regret of the former by at least 10%-50%.


Jake Feldman is an Assistant Professor at Washington University’s Olin Business School. His research interests lie at the intersection of operations management, machine learning, and algorithm design. Currently, he is primarily focused on developing models and algorithms that leverage past purchase and clickstream data to allow online retailers to make profitable product recommendation decisions.

•Abstract: Assortment Optimization for E-commerce
The first part of this talk considers a large-scale field experiment in which we compare the performance of two approaches for finding the optimal set of products to display to customers landing on Alibaba's two online marketplaces, Tmall and Taobao. Both approaches were placed online simultaneously and tested on real customers for one week. The first approach we test is Alibaba's current practice. This 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. Our second approach uses a featurized multinomial logit (MNL) model to predict purchase probabilities for each arriving customer. We use historical sales data to fit the MNL model and then, for each arriving customer, we solve the cardinality-constrained assortment optimization problem under the MNL model online to find the optimal set of products to display. Our experiments show that despite the lower prediction power of our MNL-based approach, it generates 28% higher revenue per visit compared to the current machine learning algorithm with the same set of features.

In the second part of the talk, I will introduce the click-based MNL choice model, a novel framework for capturing customer purchasing decisions in e-commerce settings. Under this model, we assume that each customer’s click behavior within product recommendation or search results pages provides an exact signal regarding the alternatives considered. We investigate the resulting assortment optimization problem, and present a case study wherein click-based MNL models and standard MNL models are fitted to historical sales and click data acquired from Alibaba. Our goal is to understand and quantify the benefits of incorporating click behavior within choice models. We demonstrate that utilizing the click-based MNL model leads to substantial improvements over the standard MNL model in term of prediction accuracy.