2016 INFORMS Annual Meeting Program

MB02

INFORMS Nashville – 2016

MB03 101C-MCC Daniel H. Wagner Prize Competition II Invited: Daniel H. Wagner Prize Competition Invited Session

3 - Response Modeling With Semi-supervised Support Vector Regression

Dongil Kim, Korea Institute of Industrial Technology, 89 Yangdaegiro-gil, Ipjang-myeon, Seobuk-gu, Cheonan, Korea, Republic of, dikim01@kitech.re.kr, Sungzoon Cho Two-stage response model has been proposed to maximize a profit of a marketing campaign by estimating the purchase amount of customers. In this paper, we propose a response modeling with Semi-Supervised Support Vector Regression (SS-SVR). In SS-SVR, label distributions of unlabeled data are estimated to consider label uncertainty. Then, training data are generated by oversampling from the unlabeled data and their estimated label distributions. Finally, a data selection algorithm is employed to reduce the training complexity. The experimental results conducted on a real-world marketing dataset showed that the proposed method improved the model accuracy and expected profit, efficiently. 4 - Support Vector Linear Regression With Multiple Instance Data Ihsan Yanikoglu, Ozyegin University, Istanbul, Turkey, ihsan.yanikoglu@ozyegin.edu.tr, Erhun Kundakcioglu We present a Support Vector Regression (SVR) framework for multiple instance (MI) data, which consists of bags of pattern vectors instead of individual instances. This setting has interesting applications such as image annotation, drug activity prediction, and causal inference over time. We provide formulations for MI regression, prove the problem is NP-hard, propose and compare efficient heuristics for the problem. MB02 101B-MCC Data Mining in Healthcare Sponsored: Data Mining Sponsored Session Chair: Ramin Moghaddass, University of Miami, 1251 Memorial Drive, MEB 308, Coral Gables, FL, 33146-0630, United States, ramin@miami.edu 1 - A Simple And Direct Projection Approach To Handling Covariate Shift Fulton Wang, MIT, fultonw@mit.edu Covariate shift is commonplace in the healthcare setting - the training population, for which labelled data is available, often differs in covariate distribution from the test population, for which predictions must be made. Covariate shift can lower test prediction accuracy even if the relation of covariates to outcomes is the same in both populations. While past methods have searched for a subspace in which the covariates of the two populations are similar, we instead propose a method that directly finds a subspace with which high test prediction accuracy can be achieved. 2 - Optimized Risk Scores In Healthcare Applications Berk Ustun, MIT, ustunb@mit.edu, Cynthia Rudin Risk scores are simple models that let users quickly assess risk by adding, subtracting, and multiplying a few small numbers. These models are widely used in healthcare, but difficult to create because they need to be risk-calibrated, use small integer coefficients, and obey operational constraints. We present a new approach to fit risk scores by solving a discrete optimization problem. We formulate the risk score problem as a MINLP, and present a cutting-plane algorithm to recover its optimal solution by solving a MIP. We use our approach to build optimized risk scores for two healthcare applications: (i) seizure prediction in the ICU; (ii) ADHD screening. 3 - Making Impact Through Identifying Impactable Members Margrét Bjarnadóttir, University of Maryland, margret@rhsmith.umd.edu A large body of research focuses on identifying patients at risk, for example for hospital readmission, appointment no-shows and declining health. However in many cases interventions to avoid adverse outcomes prove unsuccessful as patients may not be impactable, due to health status and/or the social environment. We introduce the concept of jumpers: patients at risk of adverse outcomes but who go undetected by traditional case management. We discuss the application of data mining methods to identify these members in two different settings: Diabetes management and Medicaid ED use management.

Chair: C. Allen Butler, Daniel H Wagner Associates, Inc., 2 Eaton Street, Hampton, VA, 23669, United States, Allen.Butler@va.wagner.com 1 - Data-driven Optimization For Multi-disciplinary Staffing In Mayo Clinic Improves Patient Experience Mutafa Y. Sir, Mayo Clinic, 200 First Street SW, Rochester, MN, 55905, United States, sir.mustafa@mayo.edu, David M Nestler, Thomas R. Hellmich, Devashish Das, Micheal J Laughlin, Michon Dohlman, Kalyan Pasupathy Emergency Department (ED) patient volumes fluctuate throughout the day leading to delays. Therefore, it is critical to match the staff capacity to the patient demand. A data-driven approach applied regression trees to system-generated data to produce an ideal patient volume representing ED load under optimal staffing conditions. The ideal patient volume was then used to optimize multi- disciplinary staffing levels. The new shift design significantly improved several patient-centered metrics. 2 - Optimizing New Vehicle Inventory At General Motors Robert Inman, General Motors, 30500 Mound Road, Warren, MI, 48092, United States, robert.inman@gm.com, Michael Frick, Thomas Hitchman, Robert Muiter, Jonathan Owen, Gerald Takasaki Getting inventory right enables GM to meet customer demand more efficiently. Optimizing new vehicle inventory has two dimensions: determining how many vehicles, and determining which vehicle configurations. Knowing the best aggregate number of vehicles helps manage production and pricing. Knowing the best mix of vehicles helps dealer ordering. Instead of finding “how many” to provide a given fill rate, we find the inventory that maximizes aggregate variable profit. Instead of determining “which vehicles” by simply ranking vehicle configurations by sales, we apply a practical set-covering approach to span customer demand. Topics in Power Generation Scheduling Sponsored: Energy, Natural Res & the Environment, Energy I Electricity Sponsored Session Chair: Yongpei Guan, University of Florida, 303 Weil Hall, Gainesville, FL, 32611, United States, guan@ise.ufl.edu 1 - Stochastic Scheduling For Large Scale Wind Integrated Power Systems Lixin Tang, The Institute of Industrial Engineering and Logistics Optimization, Northeastern University, Insititue of Industrial Engineering and, Logistics Optimization, Northeastern, Shenyang, 110004, China, lixintang@mail.neu.edu.cn, Jin Lang We propose a stochastic optimization problem which takes into account the volatility of large scale wind power integrated power systems. A scenario generation method which contains the information of forecast error distribution and fluctuation distribution for short-term wind power is proposed. The problem is formulated as a MINLP model. A Lagrangian relaxation algorithm is developed to solve the model. 2 - A Decomposition Approach For Hydropower Operation And Maintenance Scheduling Miguel F. Anjos, Professor and Canada Research Chair, GERAD & Polytechnique Montreal, Montreal, QC, Canada, miguel-f.anjos@polymtl.ca, Jesus A. Rodriguez, Pascal Côté, Charles Audet The generator maintenance scheduling problem is of great importance for power generation companies not only to prevent costly generator breakdowns, but also because of the impact of planned outages on the system operation. In hydroelectricity generation, the solution of this scheduling problem is complicated by uncertain water inflows, non-linear relationships between physical variables, and multiple interdependencies in space and time. Since the simultaneous solution of the hydro maintenance and operation problem is hard to achieve, we propose a linearized formulation and a decomposition approach for this problem. This work is based on the real case of Rio Tinto Aluminium in Canada. MB04 101D-MCC

148