2016 INFORMS Annual Meeting Program

TD14

INFORMS Nashville – 2016

2 - Choosing A Solution Strategy For Discrete Quadratic Optimization Robert Fourer, AMPL Optimization Inc., 4er@ampl.com The combination of integer variables with quadratic objectives and constraints is a powerful formulation tool. But when it comes to solving the resulting optimization problems, there are numerous good approaches but no one best way — even in simpler cases where the objective is convex or the constraints are linear. Both linearization of quadratic terms and quadratic generalization of linear methods turn out to be preferable in some circumstances. This presentation exhibits a variety of examples to illustrate the questions that should be asked and the decisions that must be made in choosing an effective formulation and solver. 3 - Performance Tuning For Cplex’s Spatial Branch-and-bound Solver For Global Nonconvex Mixed Integer Quadratic Programs Ed Klotz, IBM, klotz@us.ibm.com MILP solvers have been improving for more than 40 years, and performance tuning tactics regarding both adjusting solver strategies and model formulations have evolved as well. State-of-the-art global nonconvex MIQP solvers have improved dramatically in recent years, but they lack the benefit of 40 years of evolution. Also, they use a broader notion of branching that can create different performance challenges. This talk will assess the effectiveness of existing MILP tuning tactics for solving nonconvex MIQPs, as well as consider more specific strategies for this challenging type of problem. 4 - On The Benefits Of Enumeration Within An Optimization Framework Alexandra M Newman, Colorado School of Mines, Golden, CO, anewman@mines.edu While the branch-and-bound algorithm, and associated enhancements such as cuts and heuristics, vastly dominates the performance of pure enumeration for obtaining optimal solutions for (mixed) integer programming problems, this basic strategy can sometimes expedite solutions. Specifically, for cases in which enumeration of partial solutions leaves an exploitable structure in place, a small amount of enumeration over fast-solving subproblems drastically reduces solution time. We demonstrate using examples from mining (with reformulations of a monolith) and interdiction (with Benders decomposition). TD14 104D-MCC Predictive Analytics: Big Data with Purpose Sponsored: Analytics Sponsored Session Chair: Rob Lantz, Novetta Solutions, 7921 Jones Branch Drive, McLean, VA, 22102, United States, rwlantz@gmail.com 1 - Volatility-based Metrics To Analyze Network Traffic Over Time: Situational Awareness And Anomaly Detection Soumyo Moitra, Carnegie Mellon University, Software Engineering Institute, Pittsburgh, PA, 15213, United States, smoitra@sei.cmu.edu We develop some metrics to analyze temporal data that investigate different aspects of volatility. The metrics would be useful for monitoring network traffic data as well as other time series data. We discuss the motivation for the metrics and apply them to simulated data to demonstrate the properties of the metrics and to show how they can be used to derive insights into traffic patterns. Results under different scenarios are presented and compared. 2 - Security And Multidimensional Prediction Problems Anthony Boyles, Novetta, 7921 Jones Branch Drive, McLean, VA, 22102, United States, ABoyles@Novetta.com Modern security forecasting often requires comparatively high-dimensional predictions: for example, a drug bust can only be made at a specific location during the window of time that the drugs will be in that location. A forecaster must therefore be able to predict in at least three dimensions: latitude, longitude, and time. We examine techniques to reduce the complexity and increase Matthew Teschke, Novetta, 1111 Arlington Blvd., Apt. 707, Arlington, VA, 22209, United States, mteschke@novetta.com Matching, such as entity resolution, is often a required part of an analysis; however, in a Big Data environment this can be a challenging task for analysts. Naive matching implementations are not computationally feasible when record counts are measured in millions or even billions, so a different approach is needed. Blocking is a strategy for making these problems tractable, taking them from an n-squared time to near-linear time. This presentation will provide an overview of blocking, its applications, and details of implementation. predictive power of models grappling with this problem. 3 - Smart Strategies For Matching Big Data

4 - Predicting Return Abuse With Data Analytics Michael Ketzenberg, Associate Professor, Texas A&M University, Mail Stop 4217, College Station, TX, 77845, United States, mketzenberg@tamu.edu We apply data analytics to the transactional history of a large national retailer to identify the characteristics of customers who abuse return policies and then utilize this information to develop and test predictive models to help prevent such abuse. TD15 104E-MCC Joint Session AI/Analytics: Machine Learning for Public Policy Sponsored: Artificial Intelligence/Analytics Sponsored Session Chair: John J Nay, Vanderbilt University, PMB 351826 2301 Vanderbilt Place, Nashville, TN, 7235-1826, United States, john.j.nay@vanderbilt.edu 1 - Text Modeling For Understanding And Predicting The Federal Government John J Nay, Vanderbilt University, john.j.nay@vanderbilt.edu We describe the application of neural network based language modeling to better understand and predict the federal government. We embed institutions and the words from their policy documents into shared “semantic” space to explore differences across institutions with respect to policy topics. We apply our method to learn useful representations of the Supreme Court, House, Senate, and President. We also develop a machine learning approach to forecasting the probability that any bill will be enacted. The model outperforms competitor models across the three validation measures and is systematically analyzed to investigate textual and contextual factors predicting enactment. 2 - Simple Rules For Pretrial Release Decisions Jongbin Jung, Stanford University, jongbin@stanford.edu While predictive machine learning techniques might seem appealing for policy decisions, their opaque nature often render them inappropriate for the task. We investigate the possibility of constructing transparent and interpretable rules, and evaluate their performance compared to complex models in the context of pretrial release decisions. Although complex models generally outperform simple rules, we find the difference to be arguably small, especially considering the benefit of transparent rules being more likely to be implemented. 3 - Data-driven Agent-based Modeling Yevgeniy Vorobeychik, Vanderbilt University, eug.vorobey@gmail.com, Haifeng Zhang Agent-based modeling has been extensively used to study innovation diffusion. We develop a novel methodology for data-driven agent-based modeling that enables both calibration and rigorous validation at multiple scales. Our first step is to learn a model of individual agent behavior from individual event data. We then construct an agent-based simulation with the learned model embedded in artificial agents, and proceed to validate it using a holdout sequence of collective adoption decisions. We instantiate our model in the context of solar PV adoption, and utilize it to evaluate and optimize policy decisions aimed at promotion of rooftop solar. Joint Session DM/Optimization Under Uncertainty: Optimization in Data Mining and Machine Learning Sponsored: Optimization, Optimization Under Uncertainty/DM Sponsored Session Chair: Ozgu Turgut, Wayne State University, 1230 Wisteria Drive, A321, Ann Arbor, MI, 48104, United States, ozguturgut@wayne.edu Co-Chair: Michael Hahsler, Southern Methodist University, Dallas, TX, United States, mhahsler@lyle.smu.edu 1 - Sequential Aggregation-disaggregation Optimization Methods For Data Stream Mining Michael Hahsler, SMU, Dallas, TX, 75275, United States, mhahsler@lyle.smu.edu, Young Woong Park Clustering-based iterative algorithms to solve certain large optimization problems have been proposed in the past. The algorithms start by aggregating the original data, solving the problem on aggregated data, and then in subsequent steps gradually disaggregate the aggregated data. In this contribution, we investigate the application of aggregation-disaggregation on data streams, where the disaggregation steps cannot be explicitly performed on past data, but has to be performed sequentially on new data. TD16 105A-MCC

338