Informs Annual Meeting 2017

WC01

INFORMS Houston – 2017

Wednesday, 12:30 - 2:00PM

WC02

310B Decision Analysis Arcade Sponsored: Decision Analysis Sponsored Session Chair: William A. Brenneman, Procter & Gamble, 8700 Mason-Montgomery Road, Mason, OH, 45040, United States, brenneman.wa@pg.com 1 - On Data-driven Methods for Learning Graphical Models Somayeh Sojoudi, University of California-Berkeley, 1543 Delaware Street, Berkeley, CA, 94703, United States, sojoudi@berkeley.edu, Salar Fattahi Learning graphical models from data is of fundamental importance in data science, and has a wide range of applications in neuroscience, social science, biology, computer science and engineering. In this talk, we develop an optimization-based mathematical framework for three purposes: (i) comparing different data-driven methods, (ii) finding the best method from a set of candidates for a given application, and (iii) designing the best model using the selected method by optimizing the tunable parameters of the algorithm. In particular, we show that Graphical Lasso, as a popular learning technique, has a closed-form solution for sparse graphs. 2 - Comparing Machine Learning Models on Data From Wearable Sensors to Model Physical Fatigue Occurrence Zahra Sedighi, Graduate Research Assistant, Auburn University, 257 S.Gay st, apt 301B, Auburn, AL, 36830, United States, zzs0016@auburn.edu, Fadel Mounir Megahed, Lora Cavuoto In this paper, we discuss how human physical fatigue can be classified using analytical models. Best subset selection method was used to filter and identify the important features for classification. Then popular machine learning techniques were used for classification. The performance of the proposed methods was evaluated to choose the optimal classifier and prediction model. The results show that the ensemble learning outperformed the other machine learning both in prediction performance and more importantly in feature reduction. 3 - Evaluating Detection Change in a Multivariate Poisson Process Multivariate control charts have been extensively studied to signal an out-of- control process. In this study, a root transformation method on multi-attribute processes that follow a multivariate Poisson process is explored so that as to detect a change in the process using control limits. This methodology is evaluated using a simulation experiment in which a relatively stable process has a change point. Correlated quality characteristics are generated and the time of a step-change are approximated in the simulation study. The proposed method is designed to estimate the change point and find the cause for a shift in the process more effectively. 4 - Monitoring and Improving Quality via Consumer Comments William A.Brenneman, Research Fellow, Procter & Gamble, 8700 Mason-Montgomery Road, Cincinnati, OH, 45040, United States, brenneman.wa@pg.com, Alex Gutman Procter & Gamble (P&G) receives millions of consumer comments each year and requires an efficient data-mining algorithm to identify products with unexpectedly high complaint counts, as these suggest potential quality or safety issues (aka “signals”). This talk will present an overview of P&G’s adopted signal detection method, the Multi-Item Gamma Poisson Shrinker (MGPS), an empirical Bayesian method for disproportionality analysis. It will also discuss the application of text mining on consumer comments to find misclassified adverse events and identify fraudulent complaints. Hossein Najmi, University of North Texas, Denton, TX, United States, hossein.najmi@unt.edu, Robert Pavur

WC01

310A Group Decision Making Sponsored: Decision Analysis Sponsored Session Chair: Yanling Chang, Texas A&M University, 1550 Crescent Pointe Parkway, College Station, TX, 77845, United States, yanling.chang@tamu.edu 1 - Developing a Process to Rapidly Illuminate Potential Solutions Simon Goerger, Director, United States Army, 102 Windsong Cv, Ridgeland, MS, 39157-8736, United States, simon.r.goerger@usace.army.mil, Niki C. Goerger, Christina Rinaudo The SPRINTS (Systems Process for Rapid Illumination of Numerous Tentative Solutions) was developed to facilitate workshops within the Engineer Research and Development Center and with diverse groups from numerous Department of Defense organizations. This process facilitates rapid collection of stakeholder feedback to create new processes, enhance existing processes, and identify crucial capabilities requirements within an operational context. This presentation provides an overview of the SPRINTS process with particular examples from real world applications. 2 - The Effect of Shared Leadership on Team Cohesion Jie Wu, Xi’an Jiao Tong University, Xi’an, China, wujie2016@xjtu.edu.cn The existing environment with uncertainties and crises and the rise of self- management team, cross-functional team make shared leadership on the way. With the data from 54 teams composed of 284 participants analyzed by SPSS 24.0 and Mplus7.0, we explore the function of affect-and cognitive trust in team level as a bridge between shared leadership and team cohesion and discuss the different role of affect-and cognitive trust in team level between shared leadership and team output respectively. Further, shared leadership has the larger effect on team outputs in ad hoc team than established team. 3 - Voting Rules and Manipulated Evidence Youzong Xu, Xi’an Jiaotong-Liverpool University, 111 Ren’ai Road, Business Building 423B, Suzhou, 215123, China, xu.youzong@wustl.edu, Bo Li A group of voters must make a group decision based on the evidence strategically provided by a sender. The voters can choose what voting rule they use to make the group decision. We find that, if the sender can manipulate the evidence he provides, a majority rule is always superior for the voters than the unanimous rule is, because 1) the sender will provide more informative evidence under a majority rule than under the unanimous rule; 2) voters will receive greater expected utilities under a majority rule than under the unanimous rule; and 3) information equivalence can be achieved under a majority rule but not under the unanimous rule. 4 - Effectiveness-equity Tradeoffs in the Aggregation of Preferences Luis J. Novoa, The George Washington University, 2201 G Street, NW, Funger Hall, Washington, DC, 20052, United States, ljnovoa@gwu.edu, Srinivas Y. Prasad We consider the problem of aggregating preferences of a group of individuals over a set of discrete alternatives. We develop models for analyzing effectiveness equity tradeoffs using a number of different equity measures, and study how well the models conform to actual choices made by individuals in empirical studies. 5 - Value of Communication for A One-leader, Two -follower Partially Observed Markov Game Yanling Chang, Texas A&M.University, 1550 Crescent Pointe Pkwy, We consider a partially observed Markov game with three agents, a leader and two followers. We analyze how the leader’s performance is affected by changes in the quality of the communication between the two followers. Assuming each agent uses a stationary finite memory policy and that the probability of inaccurate communication between the followers is sufficiently small, we present a procedure to determine when leader may consider to modulate the communication quality between the two followers and which follower’s communication is optimal to modulate first. The intent of this research is to shed light on whether a lead agent should invest in modulating the quality of communication among followers. Apt 1206, College Station, TX, 77845, United States, yanling.chang@tamu.edu, Matthew Keblis, Ran Li

WC03A Grand Ballroom A Machine Learning and Applied Probability Sponsored: Applied Probability Sponsored Session

Chair: Daniel Russo, daniel.russo@kellogg.northwestern.edu 1 - On the Capacity of Information Processing Systems Kuang Xu, Stanford Graduate School of Business, Stanford, CA, United States, kuangxu@stanford.edu, Laurent Massoulie We study an information processing system where jobs are to be inspected by a set of experts. Inspections produce noisy results depending on the jobs’ hidden labels and the expert types, and an inspection occupies an expert for one time unit. The manager’s objective is to assign inspections so as to uncover the jobs’ hidden labels, using a minimum number of experts. Our main result is an asymptotically optimal inspection policy as the probability of error tends to zero.

504