Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB64

4 - Optimization Society’s Khachiyan Prize David Morton, Northwestern University, IEMS Department, 2145 Sheridan Road, Evanston, IL, 60208, United States Winners of the Optimization Society’s Khachiyan Prize will present their work.

n TB63 West Bldg 103B Joint Session DM/Practice Curated: Predictive Analytics and its Applications Sponsored: Data Mining Sponsored Session Chair: Talayeh Razzaghi

n TB62 West Bldg 103A

Joint Session DM/Practice Curated: Modeling and Analysis of Complex Systems with Applications Sponsored: Data Mining Sponsored Session Chair: Chun-An (Joe) Chou, Northeastern Univeristy Co-Chair: Miaolin Fan, Boston, MA, 02115, United States 1 - A Novel Framework for Multimodal Physiological Data Fusionnetwork Models of Nonlinear Dynamic Coupling Systems Miaolin Fan, Northeastern University, 360 Huntington Avenue, Boston, MA, 02115, United States The human body is considered as a complex dynamic system of multiple physiological subsystems. A novel framework was proposed to present the system as a directed network, and the interrelationship among subsystems was quantified by fusing multimodal physiological time series. Each time series is projected onto a reconstructed state space, where the temporal dependency among system’s states is captured. Then, a directed network model is formulated to characterize the coupling relationship between physiological subsystems with a temporally variable structure. We also discuss how the directed coupling can be assessed in the context of specific tasks, e.g. interpersonal communication. 2 - Cost-sensitive Feature Selection using Mixed Integer Programming Daehan Won, Binghamton University, R4, Eng. Bldg, L2, Vestal, NY, 13902, United States, Shun Cao Feature selection aims to select a subset of highly informative features that are capable of discriminating observations. Herein we consider the cost components in the selection since the traditional way may result in good selection in theory but not in practical applications. We are developing cost-sensitive classifier that minimizes the error caused by misclassification as well as maintaining the maximum amount of the cost to select the important features. To impose the cost directly, we construct two Mixed Integer Programming (MIP) models. To demonstrate the effectiveness, empirical experiments are conducted while showing that ours are capable of selecting a low-cost subset of features. 3 - A General Embedding Framework for Heterogeneous Information Learning in Large-scale Networks Na Zou, Texas A&M University, 101 Bizzell Street, 4018 Emerging Technology Building, College Station, TX, 77845, United States Network analysis has been widely applied in many real-world tasks. To extract features for these tasks, network embedding automatically learns a low- dimensional vector representation. However, it remains challenging to jointly embed the geometrical structure with heterogeneous information as well as problem of scalability. To bridge the gap, we propose a Heterogeneous Information Learning in Large-scale networks (HILL) to accelerate the joint learning. It decomposes the complex modeling into many simple and independent sub-problems. We illustrate the generalizability of HILL by applying it to perform attributed network embedding and second-order proximity learning. 4 - A Low Rank Model for Estimation of Response Function in fMRI Data Minh Pham, Rochester Institute of Technology, Rochester, NY, 22911, United States The focus of this paper is on evaluating brain responses to different stimuli and identifying brain regions with different responses using multi-subject, stimulus- evoked functional magnetic resonance imaging (fMRI) data. To jointly model many brain voxels’ responses to designed stimuli, we present a new low-rank multivariate general linear model (LRMGLM) for stimulus-evoked fMRI data. The new model not only is flexible to characterize variation in hemodynamic response functions (HRFs) across different regions and stimulus types, but also enables information “borrowing across voxels and uses much fewer parameters than typical nonparametric models for HRFs.

1 - Using Predictive Analytics to Forecast Litigation Outcomes Mohammad Javad Feizollahi, Georgia State University, 755 Ferst Drive NW, Atlanta, GA, 30332, United States, Charlotte Alexander Text mining and predictive analytics are increasingly being used to analyze and forecast the outcome of lawsuits. In this project, we parse the text of thousands of court documents filed in federal employment law cases to discover features of the plaintiffs, defendants, lawyers, and judges, and the legal claims made in each case. Together, these features help construct a model that can be deployed at different phases of litigation to predict a case’s outcome. We describe the methodology and results of this litigation prediction project. 2 - A Sensor-driven Anomaly Detection Model with a Bayesian Hierarchical Framework Ramin Moghaddass, University of Miami, McArthur Engineering Building, Coral Gables, FL, 33146, United States In this work, a new Bayesian hierarchical framework is presented that can be used to (a) model systems’ response variables in terms of system’s inputs (features) without imposing strong distributional assumptions, and (b) detect anomalies regardless of whether or not such anomalies have been observed before based on a trade-off between performance measures, such as true detection rate and false alarms. Using a Bayesian hierarchical setting, the model utilizes only a subset of important features and training samples in the training process. 3 - A Two Objective Linear Programming Approach for Data Classification Elaheh Jafarigol, University of Oklahoma, Norman, OK, 73071, United States, Theodore B. Trafalis Multi-objective optimization techniques are a useful tool for designing and analyzing supervised learning systems. This paper presents an optimization model to find support vector hyperplanes to classify large datasets with non-separable classes with modifications to the objective function in traditional support vector machine. To solve this optimization problem, parametric simplex for two- objective LPs is used. The model is implemented in Gurobi through Python to optimize the two-objective model. 4 - Predictive Analytics in Humanitarian Supply Chain using Deep Learning Donovan Fuqua, New Mexico State University, 4208 Escondido Lane, Las Cruces, NM, 88005, United States, Talayeh Razzaghi In this work, we propose the use of deep neural networks to predict shipment arrivals and system bottlenecks using multi-channel time series data. We use US Military transportation data from 2004-2015 for humanitarian relief supply chains. Although the research focuses on humanitarian operations, we will discuss multiple applications for supply chains and transportation optimization. n TB64 West Bldg 104A Joint Session DM/Practice Curated: Data Science for Block Chain, E-Business, and Commerce Sponsored: Data Mining Sponsored Session Chair: Lin Chen, University of Houston, 4800 Calhoun Road, Houston, TX, 77004, United States 1 - Effect of Mimicking News Title on Sponsored Article Engagement Quan Wang, LinkedIn, 880 W. Maude Ave, Sunnyvale, CA, 94085, United States Media companies are incorporating sponsored articles in the news feeds which used to be dedicated to editorial news articles. Using a large-scale novel dataset from a leading news website, we investigate whether mimicking news style affects the user engagement with the sponsored article. We employ a combination of human evaluation, natural language processing, binary classification models and econometrics models. We find evidence that mimicking the style of news title could lift the click probabilities by over 100% and increase conversion probability by 40%. Sub-analyses indicate that contextual congruity might be the driving force of the engagement lift.

297