2016 INFORMS Annual Meeting Program

TC01

INFORMS Nashville – 2016

Tuesday, 1:30PM - 3:00PM

Vision in humans and in non-human primates is mediated by a constellation of hierarchically organized visual areas. One important area is V4 which has highly nonlinear response properties. To better understand the filtering properties of V4 neurons we recorded from 71 well isolated cells stimulated with 4000-12000 static grayscale natural images. We fit predictive models of neuron spike rates using transformations of natural images learned by a convolutional neural network (CNN). Furthermore, we introduce new processes for interpreting such models. We conclude that the V4 neurons are tuned to a remarkable diversity of shapes such as curves, blobs, checkerboard patterns, and V1-like gratings. 2 - A New Adaptive Seizure Onset Detection Framework Sina Khanmohammadi, SUNY Binghamton, 4400 Vestal Pkwy E, SSIE Department, Binghamton, NY, 13902, United States, skhanmo1@binghamton.edu, Chun-An Chou In this study, we present a new adaptive distance-based seizure detection algorithm that provides comparable performance to more complex seizure onset detection methods in the literature using much less computational resources. The proposed framework is validated using CHB-MIT dataset, which is one of the most comprehensive scalp EEG recordings of pediatric epileptic patients. 3 - Efficient Heuristic For Large Scale Networked Data Classification Daehan Won, University of Washington, wondae@uw.edu Networked data classification is a kind of data classification where each instance is a constructed by networked structure. Similar to the current data classification, it involve huge computation time and over fitting results when the input networks have complicated structure with large size of nodes and links. To overcome those drawbacks, we present a new mathematical model based on the node selection scheme and provide a heuristic algorithm to solve the proposed math. model which is NP-hard. As a demonstration, we provide investigation results based on the human brain networks as well as simulated data set. Panel: Publication Tips Sponsored: Junior Faculty JFIG Sponsored Session Moderator: Anahita Khojandi, University of Tennessee, Knoxville, TN, United States, anahitakhojandi@gmail.com 1 - Panel Discussion: Tips For Successful Publication From Journal Editors Anahita Khojandi, University of Tennessee, Knoxville, TN, United States, khojandi@utk.edu The panelists consist of past and current editors from top journals, including Management Science, Operations Research, INFORMS Journal on Computing and Decision Analysis. The editors will share tips on how to get your paper successfully published, from selecting the right journal and preparing the manuscript, to revising the paper and responding to reviewers’ comments. They will also answer questions pertaining to writing and publication. 2 - Panelist Alice Smith, Auburn University, 3301 Shelby Center, Auburn, AL, 36849, United States, smithae@auburn.edu 3 - Panelist Jay Simon, American University, 4400 Massachusetts Avenue, NW, Washington, DC, 20016, United States, jaysimon@american.edu 4 - Panelist Alice Smith, Auburn University, Auburn, AL, United States, smithae@auburn.edu 5 - Panelist Douglas Shier, Clemson University, Clemson, SC, United States, shierd@clemson.edu 6 - Panelist Serguei Netessine, Insead, Singapore, Singapore, serguei.netessine@insead.edu TC03 101C-MCC

TC01 101A-MCC Ensemble Methods in Data Mining Sponsored: Data Mining Sponsored Session Chair: Waldyn Martinez, Miami University, 117 Country Club Dr., Oxford, OH, 45056, United States, wmartine@cba.ua.edu 1 - Ensemble Methods For Credit Risk Assessment Youqin Pan, Salem State University, Bertolon School of Business, 352 Lafayette Street, Salem, MA, 01970, United States, ypan@salemstate.edu More and more banks and financial institutions have started to pay more attention on credit risk due to the recent financial crisis. This paper aims at improving predictive powers of the credit score models using bagging and boosting algorithms. 2 - Multi-engine Out-of-sample Boosting Meinolf Sellmann, Senior Manager, IBM Research, 1101 Kitchawan Rd, Yorktown Heights, NY, 10566, United States, meinolf@us.ibm.com We present a new machine learning method that combines ensemble learning with meta-algorithmics, in particular algorithm portfolios. The result is a method that automatically determines a collection of predictive models which may or may not not consider the same concept class. To avoid over-fitting, these models are never trained on examples from regions they are later used for, nor do we ever combine predictions with each other. A portfolio method is used to select one and only one predictor at runtime, which effectively serves as regularization technique. Numerical results demonstrate that the new method massively improves the state of the art in predictive modeling. 3 - Applying Directed Acyclic Graph-based Ensemble Method For Analyzing Huge And Mixed Data In Mobile Manufacturing Seonghyeon Kang, Samsung Electrotics, Suwon, Korea, Republic of, shyeon.kang@gmail.com In mobile manufacturing, following the rapidly increasing deployment of sensing to maintain high productivity and quality, the data that we have to analyze is growing exponentially. However, in practical approach, constructing the predictive model is difficult because of the huge size of data and the mixed datatypes on the training phase. In this study, we propose the efficient ensemble method to handle those problems in mobile devices manufacturing. The effectiveness of the proposed method is demonstrated by real data from the mobile plant in one of the leading mobile companies in South Korea. 4 - Reducing The Complexity Of Ensemble Methods For Use In Large Scale Multidimensional Data Waldyn Martinez, Assistant Professor of Business Analytics, Miami University, 117 Country Club Dr., Oxford, OH, 45056, United States, martinwg@miamioh.edu Ensemble models refer to methods that combine a typically large number of fitted values into a bundled prediction. A key challenge of using ensembles in large- scale multidimensional data lies in their complexity and the computational burden they create. Recent research effort in ensembles has concentrated in reducing ensemble size, while maintaining their predictive accuracy. Here, we propose a way to reduce the complexity of an ensemble solution by optimizing on its margin distribution, while reducing their similarity. The proposed method results in an ensemble that uses only a fraction of the original weak learners, with improved or similar generalization performance. TC02 101B-MCC Data Mining in Medical and Brain Informatics I Sponsored: Data Mining Sponsored Session Chair: Chun-An Chou, SUNY Binghamton, 4400 Vestal Parkway East, Binghamton, NY, 13902, United States, cachou@binghamton.edu Co-Chair: Sina Khanmohammadi, SUNY Binghamton, 4400 Vestal Parkway East, Binghamton, NY, 13902, United States, skhanmo1@binghamton.edu 1 - Artificial Neurons Meet Real Neurons: Pattern Selectivity In V4 Reza Abbasi-Asl, University of California, Berkeley, CA, United States, abbasi@berkeley.edu, Yuansi Chen, Adam Bloniarz, Jack L. Gallant, Bin Yu

302