New Directions in
Management Science and Engineering

Data, Predictions, and Decisions

Date: Thursday, May 7, 2015
Speaker: Eric Horvitz, Director, Microsoft Research Lab at Redmond, Washington

Abstract: I will discuss directions with harnessing machine learning and inference to make predictions and to guide decision making, drawing examples from work on transportation, healthcare, and interactive systems. I will first describe efforts to build and field probabilistic models that predict flows of traffic in greater metropolitan regions. After a brief climb into the sky to discuss related efforts with air transportation, I will focus on predictive models and decision-making in healthcare. I will discuss efforts on readmissions reduction and hospital-associated infection to highlight broader possibilities with using machine learning and decision support to enhance the quality and reduce the cost of healthcare. Finally, I will discuss the use of predictive models and automated decision-making in interactive systems that leverage the complementary aspects of machine and human intelligence. I will close by highlighting several key themes rising at the intersection of machine learning and decision-making.

Bio: Eric Horvitz is a distinguished scientist and director at the Microsoft Research Lab at Redmond, Washington. His interests span theoretical and practical challenges with machine learning, inference, and decision-making. He has been elected a fellow of AAAI, ACM, the American Academy of Arts and Sciences, and the National Academy of Engineering, and has been inducted into the CHI Academy. He has served as president of the AAAI and chair of the AAAS Section on Information, Computing, and Communications, and on the advisory committees for the NSF’s Directorate for Computer & Information Science & Engineering (CISE), the Computing Community Consortium (CCC), and the DARPA Information and Science Study Group (ISAT).

Supplier Evasion of a Buyer's Audit: Implications for Motivating
Supplier Social and Environmental Responsibility

Date: Thursday, November 4, 2014
Speaker: Terry Taylor, Associate Professor, Haas School of Business, UC Berkeley

Abstract: Recently, some prominent buyers’ brands have been damaged by a supplier’ deadly factory fire or release of toxic chemicals. This paper provides guidance to buyers as to how to motivate their suppliers to exert more care to prevent such harm to workers and the environment. Obvious approaches (increasing auditing, publicizing negative audit reports, providing a loan to the supplier) can be counterproductive. Less obvious approaches (squeezing the supplier's margin by reducing the price paid to the supplier or increasing wages for workers, pre-commitment to a low level of auditing) might better motivate supplier responsibility. Even if the buyer ensures that the supplier's facility is safe, e.g., through direct investment in the facility, the supplier may outsource some production of the buyer's order to unauthorized subcontractors, exposing the buyer to risk of brand damage. The results in the paper also apply to mitigation of unauthorized subcontracting. (Joint with Erica Plambeck.)

Bio: T erry Taylor is the Milton W. Terrill Associate Professor at U.C. Berkeley’s Haas School of Business. Prior to his position at Berkeley, Terry was a professor at Columbia University’s Graduate School of Business. His current research focuses on social responsibility in operations. He is an associate editor for Management Science,Manufacturing and Service Operations Management,Operations Research, and Production and Operations Management.

A Markov Chain Approximation to Choice Modeling

Date: Thursday, October 23, 2014
Speaker: Vineet Goyal, Assistant Professor, Columbia University

Assortment planning is an important problem that arises in many industries such as retailing and airlines where the seller needs to decide on the subset of products to offer such that the expected revenues over the random preferences of the customers is maximized. The two fundamental challenges in the assortment planning problem are: i) to identify a ``good'' model for customer preferences and substitution behavior from historical sales data where preferences are not observable, and ii) to solve the resulting assortment optimization problem efficiently. Error in model selection can lead to highly sub-optimal assortment decisions. In this paper, we present a new choice model that is a simultaneous approximation for all random utility based discrete choice models including the multinomial logit, the nested logit and mixtures of multinomial logit models. Our model is based on a new primitive for substitution behavior where substitution from one product to another is modeled as a state transition of a Markov chain We show that the choice probabilities computed by our model are a good approximation to the true choice probabilities of any random utility discrete based choice model under mild conditions. In fact, the choice probabilities are exact if the underlying model is a Multinomial logit model. We also show that the assortment optimization problem under our choice model can be solved efficiently in polynomial time. Therefore, our model provides a new tractable data-driven paradigm to choice modeling and assortment optimization that is robust to model selection errors. In this talk, I will discuss properties of the new choice model and conclude with several open challenges in this area. (This is joint work with Jose Blanchet and Guillermo Gallego) The paper is available at: http://www.columbia.edu/~vg2277/MC_paper.pdf

Bio: Professor Vineet Goyal joined the Industrial Engineering and Operations Research Department in 2010. He received his Bachelor's degree in Computer Science from Indian Institute of Technology, Delhi in 2003 and his Ph.D. in Algorithms, Combinatorics and Optimization (ACO) from Carnegie Mellon University in 2008. Before coming to Columbia, he spent two years as a Postdoctoral Associate at the Operations Research Center at MIT.

Professor Goyal is interested in the design of efficient and robust data-driven algorithms for large scale dynamic optimization problems with applications in energy markets and revenue management problems. His research has been continually supported by grants from NSF and industry. He received the NSF CAREER Award in 2014 and a Google Faculty Research Award in 2013.

Optimizing a New Warehousing System

Date: Tuesday,September 30, 2014
Speaker: Zuo-Jun (Max) Shen, Chancellor’s Professor, University of California, Berkeley

Kiva Mobile Fulfillment System (MFS), developed by Kiva Systems, assists the order fulfillment process by using robots to lift and carry shelving units (i.e, inventory pods) from storage locations to picking stations, where workers pick items from the pods and put them into shipping cartons. The robots then return the empty pods to the storage area and move on with new jobs. Shen will discuss a couple of interesting problems that arise in the effort to optimize the performance of the Kiva MFS. The first is a correlated item assignment problem, where we decide which items (and in what quantities) should be put in the same inventory pod, so as to minimize the total number of visits of the pods to picking stations. An efficient algorithm is developed, and its worst-case performance bound is derived. The second is an inventory management problem, where we consider the situation where the demands of multiple products are correlated. The retailer doesn’t have the exact demand distribution, but has historic demand data as well as some prediction of future demand. The model answers two related questions: how to forecast demand for each product by utilizing the correlation information obtained from past demand data, and how to determine the optimal order quantity for each product under the newsvendor setting.

Bio: Zuo-Jun (Max) Shen is a Professor in the Department of Industrial Engineering and Operations Research, Department of Civil and Environmental Engineering at UC Berkeley and an honorary Professor at Tsinghua University. He received his Ph.D. from Northwestern University in 2000. He has been active in the following research areas: integrated supply chain design and management, market mechanism design, applied optimization, and decision making with limited information. He is currently on the editorial/advisory board for several leading journals. He received the CAREER award from National Science Foundation in 2003.

Robust Monte Carlo, Model Risk, and Counterparty Risk

Date: Thursday, April 23, 2014
Speaker: Paul Glasserman, Jack R. Anderson Professor, Columbia Business School

SPECIAL LOCATION: Math Corner 380-380Y -  4:15 to 5:15 pm

Simulation methodology has traditionally focused on measuring and reducing sampling error in simulating well-specified models; it has given less attention to quantifying the effect of model error or model uncertainty.  But simulation actually lends itself well to bounding this sort of model risk. In particular, if the set of alternative models consists of all models within a certain “distance” of a baseline model, then the potential effect of model risk can be estimated at low cost within a simulation of the baseline model. I will illustrate this approach to making Monte Carlo robust with examples from finance, where concerns about model risk have received heightened attention. The problem of bounding “wrong-way risk” in  counterparty risk presents a related question in which model uncertainty is limited to the nature of the dependence between two otherwise certain marginal models for market and credit risk.  The effect of uncertain dependence can be bounded through a convenient combination of simulation and optimization. This talk is based on work with Xingbo Xu and Linan Yang.

Bio :Professor Glasserman's research addresses risk management, derivative securities, Monte Carlo simulation, and financial stability. Prior to joining Columbia, Glasserman was with Bell Laboratories; he has also held visiting positions at Princeton University, NYU, the Federal Reserve Bank of New York, and the U.S. Treasury's Office of Financial Research. Glasserman's publications include the book Monte Carlo Methods in Financial Engineering (Springer, 2004), which received the 2006 Lanchester Prize and the 2005 I-Sim Outstanding Publication Award. Glasserman is a past recipient of the Erlang prize in applied probability (1996), an IMS Medallion (2006), and Risk Magazine's 2007 Quant of the Year Award. He was named an INFORMS Fellow in 2008.

Designing Incentive Systems for Truthful Information Sharing in Supply chains

Date: Thursday, April 25, 2013
Speaker: Ulrich Thonemann, Professor for Supply Chain Management at the University of Cologne, Director of the Department of Supply Chain Management

We consider a firm where sales is responsible for demand forecasting and operations is responsible for ordering. Sales has better information about the demand than operations and sends a non-binding demand forecast to operations. Based on the demand forecast, operations determines the order quantity. To incentivize truthful demand information sharing, we include a penalty for forecast errors in the incentive system of sales. In the utility function of sales, we also include the behavioral factors lying aversion and loss aversion. We model the setting as a signaling game and derive equilibria of the game. In a laboratory experiment, we observe human behavior that is in-line with the model predictions, but deviates substantially from expected payoff maximizing behavior. Finally, we use the behavioral model to design incentive systems for truthful information sharing and conduct an experiment to validate the approach with out-of-sample treatments and out-of-sample subjects.

Ulrich Thonemann is a Professor of Supply Chain Management at the University of Cologne. He started his academic career as an Assistant Professor of Industrial Engineering and Engineering Management at Stanford University and also holds Masters and PhD degrees from Stanford. Before joining academia, he worked as a management consultant at McKinsey & Company in Cologne. His current research focuses on Supply Chain Management, Service Management, and Behavioral Operations and on the application of state-of-the-art approaches in industry.

Cascading Processes and Network Structure

Date: Thursday, April 11, 2013
Speaker: Jon Kleinberg, Professor in Computer Science, Information Science,
Cornell University

The flow of a cascade through a network can model something desirable, such as a product or social movement that is promoted via links, or something dangerous, such as an epidemic or cascading failure that spreads contagiously. Research over the past several years has developed methods for reasoning about the connection between network structure and the ways in which cascades develop; this line of work has included probabilistic and graph-theoretic models, as well as studies of the spread of behavior in large datasets. We discuss a set of related results in this area, focusing on the distinction between "open" and "closed" network structures, and the different ways in which they can affect the outcome of a cascading process.

The talk will include joint work with Lars Backstrom, Larry Blume, David Easley, Bobby Kleinberg, Cameron Marlow, Eva Tardos, and Johan Ugander.

Systemic Risk in Networks

Date: Thursday, January 24, 2013
Speaker: Asuman Ozdaglar, Professor in Electrical Engineering and Computer Science,
Massachusetts Institute of Technology

We provide a tractable framework for studying systemic risk in networks, focusing on economic and financial networks. We first analyze the emergence of systemic risk in a production economy with an input-output structure whereby shocks to some sectors spread to their downstream sectors and beyond. We show how the nature and magnitude of systemic risk relates to the network structure of the economy, isolating the impact of first and higher order interconnections. Contrary to a common conjecture, we show that rings and other sparse regular networks are robust (i.e. as robust as a complete network) to cascades and do not generate systemic risk. Instead, systemic risk is present when some sectors are disproportionately important in the supply economy.

We then focus on interlinkages created by financial transactions (counterparty relations). We show that systemic risk in financial networks exhibits a form of phase transition as interbank connections increase. In particular, we demonstrate that as long as the magnitude and the number of negative shocks affecting financial institutions are sufficiently small, more “complete” interbank claims enhance the stability of the system. However, beyond a certain point, such interconnections start to serve as a mechanism for propagation of shocks and hence, lead to a more fragile financial system. Even in this more nonlinear financial network setting, our results thus show that the conjecture about the instability of rings is not generally true: rings are unstable when shocks are small, but it is more densely connected networks that are more unstable when shocks are large.

More Stars Stay, but the Brightest Ones Still Leave: Job Hopping in the Shadow of Patent Enforcement

Date: Thursday, December 6, 2012
Speaker: Rajshree Agarwal, Professor and Dean’s Chair in Strategy and Entrepreneurship at University of Maryland

This study investigates how a firm’s aggressiveness in enforcing patents (i.e., its “reputation for IP toughness”) affects employee-level mobility decisions. We predict and find that as a firm grows more litigious over patents, its inventive employees are less likely to ‘job hop’ to other firms within the industry. Based on evidence from the U.S. semiconductor industry, we also find that litigiousness shifts the distribution of employee exits: litigiousness operates as a sorting mechanism that aids in the retention of employees with high internal opportunities, but is less effective for employees with stronger opportunities for outside advancement. The study provides the first systematic evidence that the mobility of knowledge workers is shaped not only by state laws on non-compete agreements, but also by firm-specific reputations built through patent enforcement.

The Genomics Revolution

Date: Thursday, November 15, 2012
Speaker: David Heckerman, Senior Director, eScience Research Group, Microsoft Research

A little over a decade ago, the first human genome was sequenced at a cost of about one-hundred million dollars. Today, it costs well under ten thousand dollars. In fact, the cost is dropping much more quickly than the rate of Moore’s law. As a result, genomics data is becoming widely available and widely used. I will discuss elements of this genomics revolution relevant to Management Science and Engineering, including probabilistic modeling for the identification of genetic causes of disease. I will also touch on legal and ethical issues surrounding this revolution.

On Stochastic Insurance and Reinsurance Risk Networks

Date: Thursday, November 8, 2012
Speaker: Jose Blanchet, Associate Professor of Engineering, Industrial Engineering & Operations Research, Columbia University

In the last few years, substantial interest has been given to the study of systemic risk. In this talk we describe a class of models for systemic risk analysis in the setting of insurance and reinsurance participants. The models are constructed with the aim of capturing features such as cascading effects at the time of default due to counter-party risk and contagion. We also impose a probabilistic structure that allows us to rigorously study risk analysis questions using the theory of combinatorial optimization. In the end, we are interested in answering questions such as: a) What group of companies are the most relevant from a systemic risk standpoint? a) How do we quantify the role of reinsurance companies in systemic risk? c) How do we to understand and quantify the role of a regulator and associated capital requirements for systemic risk mitigation?/p>

A General Framework for Systemic Risk

Date: Tuesday, October 23, 2012
Speaker: Ciamac Moallemi, Associate Professor in the Decision, Risk, & Operations Division of the Graduate School of Business at Columbia University.

Systemic risk [SR] refers to the risk of collapse of an entire complex system, as a result of the actions taken by the individual component entities or agents that comprise the system. SR is an issue of great concern in modern financial markets as well as, more broadly, in the management of complex business and engineering systems. We propose an axiomatic framework for the measurement and management of SR based on the simultaneous analysis of outcomes across agents in the system and over scenarios of nature. Our framework defines a broad class of SR measures that accommodate a rich set of regulatory preferences. This general class of SR measures captures many specific measures of SR that have recently been proposed as special cases, and highlights their implicit assumptions. Moreover, the SR measures that satisfy our conditions yield decentralized decompositions, i.e., the SR can be decomposed into risk due to individual agents. Furthermore, one can associate a shadow price for SR to each agent that correctly accounts for the externalities of the agent’s individual decision-making on the entire system. This is joint work with Chen Chen and Garud Iyengar

Content Referral on the Internet

Date: Thursday, October 11, 2012
Speaker: Assaf Zeevi, Henry Kravis Professor at the Graduate School of Business, Columbia University

Search engines have been a transformative force, shaping the way we seek content and navigate the world wide web. They have also provided one of the most successful and robust revenue models on the Internet in the form of sponsored search. Recent trends in online user behavior exhibit interesting new phenomena, one of which is the increasing prevalence of "wandering'' type patterns; seeking content, yet in a less targeted manner. What are these users looking for? What current content is potentially attractive? And finally, how does one match the two in real time? This talk will explore some of these questions, with the intent of illustrating the use of data to inform modeling, and some of the analytical challenges that ensue.

Sparse Machine Learning for Large Text Corpora

Date: Tuesday, May 15
Speaker: Laurent El Ghaoui, Professor of EECS and IEOR at UC Berkeley

Sparse machine learning has recently emerged as powerful tool to obtain models of high-dimensional data with high degree of interpretability, at low computational cost. These methods can be extremely useful for understanding large collections of text documents, without requiring user expertise in machine learning. Our approach relies on three main ingredients: (a) multi-document text summarization and (b) comparative summarization of two corpora, both using sparse regression or classification; (c) sparse principal components and sparse graphical models for unsupervised analysis and visualization of large text corpora. In this talk I will describe some theoretical and algorithmic challenges related to these approaches, as well as experiments pertaining to real-life text databases, from scientific literature (PubMed), news and internal industry reports (commercial pilots reports).

Adventures in Policy Modeling!

Date: Tuesday, April 17
Speaker: Ed Kaplan, William N. and Marie A. Beach Professor of Mgmt Sciences, ChE & EnvE Engineering & Public Health, Yale University

Policy Modeling refers to the application of operations research, statistics, and other quantitative methods to model policy problems. Recognizing that analyses of all sorts often exhibit diminishing returns in insight to effort, the hope is to capture key features of various policy issues with relatively simple “first-strike” models. Problem selection and formulation thus compete with the mathematics of solution methods in determining successful applications: where do good problems come from? how can analysts tell if a particular issue is worth pursuing? In addressing these questions, I will review some personal adventures in policy modeling selected from public housing, HIV/AIDS prevention, bioterror preparedness, suicide bombings and counterterrorism, in vitro fertilization, predicting presidential elections, and March Madness.

Engineering Happiness

Date: Thursday, April 12
Speaker: Rakesh Sarin, Paine Professor of Management, UCLA Anderson School

We develop and apply a unique and novel application of analytics and Decision Analysis to the study of happiness. Our results should be useful to individuals seeking to become happier and to organizations that wish to improve customer and employee satisfaction and productivity. Our model begins with the fundamental equation: HAPPINESS = REALITY – EXPECTATIONS. Following this, we propose a set of six laws that modify the fundamental equation, making it more precise and applicable to a wide range of situations and choices.