Informs Annual Meeting 2017

TA38

INFORMS Houston – 2017

TA38

that violate the problem context and requirements. Since the k-means algorithm is so intuitive, we modify how the algorithm forms clusters so that cluster formations tend to have a similar number of observations in each group. We apply our algorithm to an industry problem posed by division III men’s wresting where they were seeking to form new conference regions that we more competitively and geospatially balanced. We show our algorithm improves their current region assignments. Lastly, we identify additional potential avenues this clustering algorithm could be implemented.

352C Multi-Criteria Decision Making: Can we Always Know that we have Made the Right Decision? Sponsored: Multiple Criteria Decision Making Sponsored Session Chair: Evangelos Triantaphyllou, Louisiana State University, Baton Rouge, LA, 70803, United States, etriantaphyllou@yahoo.com Co-Chair: Juri Yanase, Complete Decisions, LLC, Baton Rouge, LA, 70810, United States, jurijuriy@aol.com 1 - How to (and not) Apply Multi-Criteria Decision Analysis (MCDA) to Medical Decisions Edouard Kujawski, 1159 Miller Ave., Berkeley, CA, 94708, United States, edkuj@comcast.net MCDA is increasingly being implemented in shared medical decision-making. Medical professionals and patients face a quagmire of deceitfully attractive methods that can substantially impact alternative rankings. Medical decision- making is often a problem under uncertainty: tests provide imperfect information and outcomes are probabilistic. A significant effort is required to develop realistic and mathematically valid models for making reliable life-critical medical decisions. 2 - Some Important Issues Regarding the use of a Number of Well-known MCDM Methods Fujun Hou, Beijing Institute of Technology, No.5 South Street Zhongguancun, Haidian, Beijing, 100081, China, houfj@bit.edu.cn Recent findings on the way some popular MCDM methods rank and prioritize alternatives, raise questions on the validity of their results. We examine these issues comprehensibly. We consider three types of preference data for MCDM problems and identify some isomorphic relations. We use these results to design some appealing approaches for dealing with the problems that face current MCDM methods. 3 - Can We Always Determine the Best MCDM Method? Evangelos Triantaphyllou, Louisiana State University, Dept of Computer Science, 298 Coates Hall, Baton Rouge, LA, 70803, United States, etriantaphyllou@yahoo.com, Juri Yanase The reason there is more than one method to solve a given MCDM problem is the fact that sometimes such methods may yield different results for exactly the same input data. Thus, a critical question is to identify the best MCDM method. This problem may be intractable. We discuss a number of intriguing issues related to this fundamental MCDM question. Sponsored: Analytics Sponsored Session Chair: Aaron Burciaga, Accenture, Global Analytics Executive, 4305 Majestic Lane, Fairfax, VA, 22033, United States, adburciaga@gmail.com 1 - A Decision Support Prototype for Blood Donation Management Deepti Bahel, Proxima Analytics, West Lafayette, IN, 47906, United States, baheldeepti@gmail.com, Matthew Lanham We study the performance of machine learning algorithms that have not been previously investigated to support the problem of blood donation prediction. We build models on clustered data sets using k-means clustering and not using clustering to see if performance can be significantly improved. The motivation for this research is that blood demand is gradually increasing by the day due to needed transfusions due to accidents, surgeries, diseases etc. Accurate prediction of the number of blood donors can help medical professionals know the future supply of blood and plan accordingly to entice voluntary blood donors to meet demand. In addition to predicting demand accurately, we show how to incorporate the costs of targeting a potential donor or not, which provides novel decision support to blood centers in how they can manage their donation efforts. 2 - A Balanced k-Means Clustering Algorithm to Support the Division TA39 352D Young Researchers

TA41

352F Health Care, Modeling and Optimization Contributed Session Chair: Aditya Jain, Baruch College, Zicklin School of Business, New York, NY, United States, aditya.jain@baruch.cuny.edu 1 - Evaluating Cardiac Mortality Rates: An Alternative to the RAMR Christine Pitocco, Research Professor, Stony Brook University, 333 Harriman Hall, Stony Brook, NY, 11794-3775, United States, christine.pitocco@stonybrook.edu, Acsah Joseph, Thomas Sexton The risk adjusted mortality rate (RAMR) measures the mortality performance of health care providers such as hospitals and physicians. It is used by health care consumers in selecting providers and policy makers evaluating their mortality performance. However, the RAMR has several serious flaws. We propose an alternative measure and demonstrate its use in measuring performance of cardiac surgeons and hospitals. 2 - Graph Regularized EEG Inverse Program with Total Variation Priors Feng Liu, UT.Arlington, 19251 Preston Rd, Apt 1214, Dallas, TX, 75252, United States, liufengchaos@gmail.com As EEG sensors measure electrical potentials on the scalp instead of directly measuring activities of brain sources, many approaches were proposed to infer brain source activation pattern. We propose that by leveraging label information, the task related discriminative sources can be much better retrieved among strong spontaneous background signals. A novel supervised EEG source imaging model was proposed which aims to explicitly extract the discriminative source extents by implicitly coding the label information into the graph regularization term. An ADMM optimization algorithm is given to solve the problem. Numerical results show the effectiveness of our proposed framework. 3 - Modeling the Pre-clinical Phase for Ovarian Cancer using a Conditionally Stationary Markov Process Jing Voon Chen, University of Southern California, Los Angeles, CA, 90089, United States, jingvooc@usc.edu, Julia L. Higle Creating a nonstationary natural history (NH) model can overwhelm parameter search methods and exacerbate the problems surrounding calibration uncertainty, especially when data is scarce. We present a conditionally stationary Markov model (CSMM) that decomposes the overall NH into two components: a stationary Markov chain governing the disease progression and a nonstationary process for competing risk mortality. We investigate the window of opportunity for ovarian cancer using a CSMM. 4 - Creating Thickness in Kidney Exchange Markets We study a dynamic model of kidney exchange in which patient-donor pairs arrive and depart continuously and the composition of the market depends on the matching algorithm. We discuss the role of thickness and information in designing optimal policies. 5 - The Effect of Flexible Service Rates in Appointment Scheduling System Aditya Jain, Baruch College, Zicklin School of Business, 55 Lexington Ave, Suite 9-240, New York, NY, 10010, United States, aditya.jain@baruch.cuny.edu, William P.Millhiser Recent empirical evidence suggests that ambulatory care healthcare workers intentionally vary service rates in response to two factors: time of day and queue length, where workers choose work speeds to balance their own overtime and the patient’s wait. We explore a theoretical justification of this behavior and show that workers speed up if the cost of patient waiting is relatively small compared to workers’ overtime cost. We analyze the implications of such worker behavior on various performance metrics for established appointment systems, and how appointment systems should be adapted. Shayan Oveis Gharan, Stanford University, 475 Via Ortega,, Huang Engineering Center, 314K, Stanford, CA, 94305, United States, shayan@stanford.edu, Shengwu Li

III Men’s Wresting Conference Alignment Problem Amar Iyengar, Purdue University, West Lafayette, IN, United States, iyengara@purdue.edu, Matthew Lanham

We develop a balanced k-means clustering algorithm that allows the modeler to specify a range of how many observations the modeler would like within each generated cluster. The traditional k-means clustering algorithm is one of the most well-known clustering algorithms used by practitioners today. Unfortunately, it often leads to the generation of groups that have a diverse number of observations in each group. Such scenarios regularly lead to cluster formations

284