2015 Informs Annual Meeting

SB32

INFORMS Philadelphia – 2015

2 - Reliability of a Two Parallel K Out of N System with Removable Repair Mechanism George Mytalas, S. Lecturer, NJIT, University Heights, Newark, NJ, United States of America, mytalas@njit.edu We study the reliability of a (k1,k2)-out-of-(n1,n2) system which consists of two different types of components with finite populations and a single repair machine. The system operates under the (N1,N2)-policy i.e. the server is activated for exhaustive repairs. The repaired components are assumed to be as good as new. Repair times of components and life times are assumed to be independent of each other. 3 - Bullwhip Effect in a Pharmaceutical Supply Chain Ming Jin, University of Utah, 1655 E Campus Center Dr, Salt Lake City, UT, United States of America, ming.jin@utah.edu, Glen Schmidt, Nicole Dehoratius We investigate the bullwhip effect in a pharmaceutical supply chain. Specifically, we estimate the bullwhip effect at the stock keeping unit (SKU) level, analyze the bias in aggregated measurement of the bullwhip effect, and examine various driving factors of the bullwhip effect. Data aggregation across products or over long time periods tends to mask the bullwhip effect in some cases. Price promotion, order batching, and inventory are three main factors associated with the bullwhip effect. Sponsor: Analytics Sponsored Session Chair: Drew Pulvermacher, Director, Decision Sciences, PerformanceG2, Inc., 3432 Sunset Dr, Madison, 53705, United States of America, drew@performanceG2.com 1 - Descriptive, Predictive and Prescriptive Analysis for and by Business Users Alain Chabrier, IBM Spain, Santa Hortensia 26-28, Madrid, Spain, achabrier@es.ibm.com, Xavier Ceugniet, Stéphane Michel Visual analytics tools such as Watson Analytics provide easy ways for business users to benefit of descriptive and predictive analytics. Using configuration and elicitation of constraints and goals based on suggestions and natural language, we show how prescriptive analytics can also be supported and more complex business problems solved using the combinations of the 3 analytics area. We illustrate with a campaign marketing optimization use case. 2 - The Five Minute Analyst Harrison Schramm, Navy Headquarters Staff, 1507 22nd Street South, Arlington, VA, 22202, United States of America, Harrison.Schramm@gmail.com A little bit of analysis can go a long way – provided it’s both scoped appropriately and presented in a way that is appealing to the layperson. The Five Minute Analyst, appearing in INFORMS/Analytics Magazine, attempts to apply some analysis to everyday problems with the goal of both improving decisions and sharing the power of a little bit – five minutes’ worth, to be precise – of analysis towards everyday problems. This session will be a ‘best of’, going further in depth on the articles I’ve written, as well as discussing some ‘also rans’ – interesting problems that did not make the column. The application areas contain (but are not limited to) probability, statistics and game theory, although others may find their way in. Problems addressed will include Star Wars and Downton Abbey. There will be zombies. SB31 31-Room 408, Marriott Predictive Analytics for Health Care Decision Making Chair: Nick Street, Professor And Departmental Executive Officer, The University of Iowa, 108 Pappajohn Business Building, S210, Iowa City, IA, 52242, United States of America, nick-street@uiowa.edu 1 - Sparse Logical Machine Learning Models Cynthia Rudin, Associate Professor, Massachusetts Institute of Technology, MIT, Cambridge, MA, United States of America, rudin@mit.edu CART (Classification and Regression Trees) is possibly the most popular machine learning method in industry. For the last few years, my research group has been trying to build competitors for CART. I will overview our work on sparse logical models that are direct competitors for CART. They are competitive in all ways: accuracy, interpretability, and tractability. SB30 30-Room 407, Marriott Model Building like a Boss Sponsor: Data Mining Sponsored Session

2 - Using Machine Learning Approaches to Predict Patient Risk from EHR Data Alexander Cobian, Department of Computer Sciences, University of Wisconsin-Madison, 1300 University Ave, Madison, WI, 53706-1532, United States of America, cobian@cs.wisc.edu, Mark Craven We explore the task of learning from electronic health record (EHR) data to predict patient risk levels for conditions of interest (asthma exacerbation and post-operative deep venous thrombosis.) While standard risk questionnaires focus on small numbers of known risk factors, we consider thousands of variables elicited from the EHR in order to identify novel risk indicators. Further, our approach attempts to discover latent variables that connect related, observed variables. 3 - Reducing Patient Risk through Inverse Classification: An SVM-based Method Michael Lash, The University of Iowa, 318 MacLean Hall, In this work we propose a novel algorithm to address the problem of inverse classification, and apply the result to a recommendation system for patient risk minimization. We propose a mixed-integer nonlinear programming based on SVMs that finds the optimal values of the attributes that achieve the targeted probability of being in a desired class. The result is a flexible model that arrives at a set of realistic recommendations to mitigate patient risk. 4 - Understanding Emergency Department 72-hour Revisits Among Medicaid Patients Kristin Bennett, Mathematical Sciences Dept., Rensselaer Polytechnic Institute, 110 8th Street, AE 327, Troy, NY, 12180, bennek@rpi.edu, James Hendler, James Ryan We analyze emergency department (ED) usage at one hospital to understand ED return visits within 72 hours. “Frequent flier” patients with multiple revisits account for 47 percent of Medicaid patient revisits over a two year period. Statistical and L1-logistic regression analysis reveals distinct patterns of ED usage between frequent- and infrequent-patient encounters suggesting distinct opportunities for interventions to improve care and streamline ED workflow. SB32 32-Room 409, Marriott Advances in Community Detection and Influence Analysis in Social Networks Sponsor: Data Mining Sponsored Session Chair: Wenjun Wang, University of Iowa, S283 Pappajohn Business Building, Iowa City, IA, 52242, United States of America, wenjun-wang@uiowa.edu 1 - Finding Hierarchical Communities in Complex Networks Wenjun Wang, University of Iowa, S283 Pappajohn Business Building, Iowa City, IA, 52242, United States of America, wenjun-wang@uiowa.edu, Nick Street Based on a Shared-Influence-Neighbor (SIN) similarity measure, we propose two novel influence-guided label propagation (IGLP) algorithms for finding hierarchical communities in complex networks. One is IGLP-Weighted-Ensemble (IGLP-WE), and the other is IGLP-Direct-Passing (IGLP-DP). Extensive tests demonstrate superior performance of our methodology in terms of excellent quality and high efficiency in both undirected/directed and unweighted/weighted networks. 2 - Community Detection in Dynamic Networks with Multiple Attributes Xiang Li, University of Florida, Room 555, CSE Building U. of Florida, Gainesville, FL, 32611, United States of America, xixiang@cise.ufl.edu, My Thai In this talk, we provide a mathematical model to qualify the relationship between communities based on network structure and that based on the common node attributes. We next present a dynamic and scalable algorithm to detect such communities, based on the multiplex graph theory. Iowa City, IA, 52242, United States of America, michael-lash@uiowa.edu, Nick Street, Qihang Lin

75