Informs Annual Meeting 2017

MD29

INFORMS Houston – 2017

4 - On Finding Stable and Efficient Solutions for the Constrained Coalition Formation Problem Hoda Atef Yekta, PhD Candidate, University of Connecticut, 2100 Hillside Rd,, Storrs, CT, 06869, United States, hoda.atefyekta@business.uconn.edu, David Bergman, Robert Day This research explores the constrained coalition formation problem with constraints on the size of each coalition and the characteristics of the members of each coalition. We develop a branch-and-cut-and-price (BCP) algorithm to address both minimizing the maximum instability, and maximizing total utility, simultaneously. This problem arises in a variety of real-world settings, including project management, military platooning, and peer-to-peer ride-sharing 350E Making Computationally Difficult Decisions Sponsored: Artificial Intelligence Sponsored Session Chair: Meisam Razaviyayn, Ph.D., University of Southern California, 3715 McClintock Ave, Los Angeles, CA, 90089, United States, razaviya@usc.edu 1 - Stochastic Primal-Dual Methods and Sample Complexity of Markov Decision Processes Yichen Chen, Graduate Student, Princeton University, 751 Hibben Magie Road, Apt 109, Princeton, NJ, United States, yichenc@princeton.edu We study the online estimation of the optimal policy of a Markov decision process (MDP). We propose a class of Stochastic Primal-Dual (SPD) methods which exploit the inherent minimax duality of Bellman equations. The SPD methods update a few coordinates of the value and policy estimates as a new state transition is observed. We first consider a basic version of SPD that uses Euclidean projection for both the primal and dual updates. We then propose an accelerated version of SDP that uses relative entropy projection in the dual update. For MDPs that are ``sufficiently” ergodic, the improved SPD has a near-optimal sample/running-time complexity. 2 - Nonnegative Polynomials, Nonconvex Polynomial Optimization, and Applications to Learning Georgina Hall, Princeton University, Charlton Street, Sherrerd Hall, Princeton, NJ, 08544, United States, gh4@princeton.edu We give a brief overview of the developments in the field of sum of squares optimization and show how it can be applied to two problems at the interface of machine learning and polynomial optimization. In part (i), we study the problem of learning a monotone polynomial from data. This is motivated by regression problems where the underlying function to be learned is monotone (consider the price of a car as a function of its fuel efficiency). In part (ii), we study the problem of optimally decomposing a multivariate polynomials as the difference of two convex polynomials. This is motivated by certain majorization-minimization algorithms used in nonconvex optimization that require such a decomposition. 3 - Learning to Optimize: Training Deep Neural Networks for Wireless Resource Management Mingyi Hong, University of Virginia, 3019 Black Engineering Building, IMSE Department, Ames, IA, 50011, United States, mingyi@iastate.edu We propose a new learning-based approach for wireless resource management. The idea is to treat the input and output of a resource allocation algorithm as an unknown non-linear mapping and use a deep neural network (DNN) to approximate it. If the non-linear mapping can be learned accurately and effectively by a DNN of moderate size, then such DNN can be used for resource allocation in almost real time, since passing the input through a DNN to get the output only requires a small number of simple operations. We first characterize a class of ‘learnable algorithms’ then design DNNs to approximate some algorithms of interest in wireless communications. We use extensive numerical simulations to demonstrate the superior ability of DNNs for approximating two considerably complex algorithms that are designed for power allocation in wireless transmit signal design. 4 - From Predictive Methods to Missing Data Imputation: An Optimization Approach MD29

state-of-the-art methods. We also discuss extensions to explicitly account for multi-class categorical variables.

MD30

350F Topics in Text Analytics Invited: Social Media Analytics Invited Session Chair: Zachary Davis, Virginia Tech, 1007 Pamplin Hall, Blacksburg, VA, 24061, United States, zached1@vt.edu Co-Chair: Alan Abrahams, Virginia Institute of Technology, Blacksburg, VA, 24061, United States, abra@vt.edu 1 - How to Make a Persuasive Pitch: An Extended Text-mining Model for Crowdfunding Analysis Sukhwa Hong, Virginia Tech, Pamplin College of Business, 880 West Campus Drive, Blacksburg, VA, 24061, United States, sukhwa@vt.edu, David Michael Goldberg, Onur Seref, Christopher Zobel Crowdfunding is a new funding source and growing in popularity, providing a vital service to individuals especially in developing countries, who could not otherwise get the financial resources they need. Understanding factors that determine the funders’ decisions is important in writing persuasive crowdfunding pitches for successful projects. In this paper, we present an integrated text-mining model that combines the Bag-of-Words model and related variables for predicting the success of crowdfunding projects in Kiva. Furthermore, association rule mining is used to identify factors that influence the funders’ decisions for writing persuasive crowdfunding pitches for successful projects. 3 - Temporal Effects, Opinion Extremity, and Information Quality in Online Word of Mouth David Michael Goldberg, Virginia Tech, Blacksburg, VA, 24061, United States, goldberg@vt.edu Online text frequently describes increasingly extreme opinions over time as additional contributors seek to differentiate their opinions from the crowd. We examine how information quality is affected by this phenomenon, as these extreme viewpoints may not reflect the general consensus. 4 - Impact of Health Social Media on the Health State of Patients with Chronic Illness Zachary Davis, Virginia Tech, 1007 Pamplin Hall, Blacksburg, VA, Online health communities are utilized by a vast number of people with limited objective evidence regarding the impact on their actual health state. We examine an online health community consisting of members managing a chronic illness and empirically analyze objective and fully observable data regarding their health state to determine if participation and the giving and receiving of social support has an impact on the overall health state. Our novel contribution is the identification of the observable measure of the health state and our preliminary findings indicate that the use of health social media has a positive impact on the health state. 351A Social Media Analytics Invited: Social Media Analytics Invited Session Chair: Fujie Jin, jinfujie@wharton.upenn.edu 1 - Online vs. Offline: A Structural Estimation of Consumer Purchase Behavior Luna Zhang, PhD, Lehigh University, College of Business and Economics, xiz313@lehigh.edu, Daisy Dai, Oliver Yao Using large-scale POS datasets, we develop a structural model and estimate the demand for consumer packaged goods in both online and offline channels. Based on our structural model estimates, we conduct counterfactual analyses and examine consumer purchase behavior between the focal product and substitute products when the price changes and stockout occurs. We find that: 1) online own-price elasticity is 1.41 times higher than offline own-price elasticity 2) offline cross brand cross-price elasticity is 2.49 times higher than that online; 3) offline consumers are 3.96 times more likely to purchase substitute brands when the focal brand is out-of-stock as compared with online consumers. 24061, United States, zached1@vt.edu, Qianzhou Du, Alan Gang Wang, Christopher Zobel, Lara Z Khansa MD31

Colin F. Pawlowski, Massachusetts Institute of Technology, Cambridge, MA, 02143, United States, cpawlows@mit.edu, Dimitris Bertsimas, Ying Zhuo

We propose a framework based on optimization to impute missing data which can readily incorporate predictive models including K-NN, SVM, and decision tree based methods. We derive a family of fast first-order methods opt.impute that obtains high quality solutions in seconds for data sets with 100,000s of missing values. In large-scale synthetic and real data experiments, opt.impute produces the best overall imputation in almost 80% of all data sets benchmarked against

246