2015 Informs Annual Meeting

SA20

INFORMS Philadelphia – 2015

4 - Social Structure Optimization in Nurse Scheduling Problem Alireza Farasat, Graduate Research Assistant, University at Buffalo (SUNY), 327 Bell Hall, Department of Industrial and Systems Eng, Amherst, NY, 14260, United States of America, afarasat@buffalo.edu, Alexander Nikolaev This paper presents a mathematical framework for treating the Nurse Scheduling Problem (NSP) explicitly incorporating Social Structure (NSP-SS). While traditional approaches generate a configuration of individual schedules, the presented framework introduces models that assign nurses to working shifts to achieve an optimal structure of individual attributes and social relations within the teams. For an NP-Hard instance of NSP-SS, an integer program is presented, followed by a LK-NSP heuristic. SA18 18-Franklin 8, Marriott Recent Advances on Support Vector Machines Research Cluster: Modeling and Methodologies in Big Data Invited Session Chair: Shouyi Wang, Assistant Professor, University of Texas at Arlington, 3105 Birch Ave, Grapevine, TX, 76051, United States of America, shouyiw@uta.edu 1 - Fast Scalable Support Vector Machines for Big Bimodical Data Analytics Talayeh Razzaghi, Postdoctoral Research Fellow, Clemson University, 221 McAdams Hall, Clemson University, Clemson, United States of America, trazzag@clemson.edu, Ilya Safro, Mark Wess Solving the optimization model of support vector machines is often an expensive computational task for very large biomedical training sets. We propose an efficient, effective, multilevel algorithmic framework that scales to very large data sets. Our multilevel framework substantially improves the computational time without loosing the quality of classifiers for balanced and imbalanced datasets. 2 - Value-at-Risk Support Vector Machine (Var-SVM ): MIP Representation and Equivalence of Formulations Victoria Zdanovskaya, Research And Teaching Assistant At Industrial And Systems Engineering Department, University of Florida, 303 Weil Hall, Gainesville, FL, 32611, United States of America, ladyvi@ufl.edu, Konstantin Pavlikov SVMs is a widely used data classification technique. A class of Var-SVMs is known to be robust to the outliers in the training dataset. Unfortunately Var-SVM is a nonconvex optimization problem. We consider MIP representations of Var-SVM, that can be solved by standard Branch & Bound algorithm. We also consider different techniques that help to dramatically improve computational performance of such formulations. 3 - A Comparison of Constraint Relaxation and Bagging Policies in Support Vector Classification In classification, when data are available in uneven proportions the problem becomes imbalanced and the performance of standard methods deteriorates. Imbalanced classification becomes a more challenging in the presence of outliers. In this presentation, we study several algorithmic modifications of support vector machines for such problems. We show that the combined used of cost sensitive learning with constraint relaxation performs better compared to approaches that involve bagging. 4 - Semi-supervised Proximal Support Vector Machine with Sparse Representation Regularization Jiaxing Pi, University of Florida, 3800 SW 34th St. Apt. P138, Gainesville, FL, 32608, United States of America, jiaxing@ufl.edu, Panos Pardalos Proximal Support Vector Machine has been an efficient technique to generate classifiers. Sparse representation can detect neighborhood for a signal by reconstructing it with the linear span of other data. We applied sparse representation to build a regularization which can achieved semi-supervised assumptions for unlabeled data. Experiment on standard datasets are performed to compare the proposed framework with PSVM with manifold regularization. Petros Xanthopoulos, University of Central Florida, 12800 Pegasus Dr., Orlando, FL, 32816, United States of America, petrosx@ucf.edu, Onur Seref, Talayeh Razzaghi

5 - Extending Relaxed Support Vector Machines Orestis Panagopoulos, University of Central Florida, 12800 Pegasus Dr., Orlando, FL, 32816, United States of America, opanagopoulos@knights.ucf.edu, Onur Seref, Talayeh Razzaghi, Petros Xanthopoulos In this work, we propose Relaxed Support Vector Regression (RSVR) and One- Class Relaxed Support Vector Machines (ORSVM). The methods constitute extensions of Relaxed Support Vector Machines (RSVM). They are formulated using both linear and quadratic loss functions and are solved with sequential minimal optimization. Numerical experiments on public datasets and computational comparisons with other popular classifiers depict the behavior of our proposed methods. SA19 19-Franklin 9, Marriott High-performace Computation for Optimization Sponsor: Computing Society Sponsored Session Chair: Suresh Bolusani, Lehigh University, 524 Montclair Avenue, Bethlehem, United States of America, sub214@lehigh.edu 1 - Distributed Integer Programming Ezgi Karabulut, Georgia Institute of Technology, 755 Ferst Drive, NW, Atlanta, GA, 30332-0205, United States of America, ezgi.karabulut@gatech.edu, George L. Nemhauser, Shabbir Ahmed We want to find distributed solution algorithms for integer programming problems that allow only minimal interaction between the solvers. 2 - Scalable Communication in Parallel Optimizaiton Oleg Shylo, University of Tennessee, 851 Neyland Drive, 523 John Tickle Building, Knoxville, TN, United States of America, oshylo@utk.edu We establish theoretical models of algorithm portfolios to optimize communication patterns in algorithms, closely match empirical behavior of communicative algorithm portfolios, and predict computational performance for new and untested configurations. 3 - Solving Bilevel Linear Optimization Problems in Parallel Suresh Bolusani, Lehigh University, 524 Montclair Avenue, Bethlehem, PA, United States of America, sub214@lehigh.edu, Ted Ralphs Many real world applications involve multiple, independent decision makers with multiple, possibly conflicting objectives. Bilevel linear optimization provides a framework for modeling of such problems. With the growing number of applications, faster solution algorithms for bilevel optimization problems are needed. In this work, we present a parallel approach to solving bilevel optimization problems. Computational results will be presented. Big Data in the Clouds Cluster: Cloud Computing Invited Session Chair: Lydia Chen, IBM Zurich, yic@zurich.ibm.com 1 - Declarative Cloud Performance Analytics Boon Thau Loo, Associate Professor, University of Pennsylvania, Philadelphia, PA, 19104, United States of America, boonloo@cis.upenn.edu This talk presents Scalanytics, a declarative platform that supports high- performance cloud application performance monitoring. Scalanytics uses stateful network packet processing techniques for extracting application-layer data from network packets, a declarative rule-based language for compactly specifying analysis pipelines, and a parallel architecture for processing network packets at high throughput. I will next describe the commercialization of Scanalytics as Gencore (gencore.io). SA20 20-Franklin 10, Marriott

43