Informs Annual Meeting 2017

TA71

INFORMS Houston – 2017

3 - A Semi-markov Mixture Model Clustering Method for Multi-state Partially Censored Data Geet Lahoti, Graduate Research Assistant, Georgia Tech, 1038 Mcmillan St. NW, Atlanta, GA, 30318, United States, glahoti6@gatech.edu, Chitta Ranjan, Zih Huei Wang, Zhen Qian, Chuck Zhang, Ben Wang Semi-Markov model is used to model the state transition dynamics in spatio- temporal data widely found in the web, marketing, reliability, and healthcare. The main challenge in modeling this data is the underlying (unobserved) heterogeneity with respect to the behaviors in the transition dynamics. We propose a semi-Markov mixture model based methodology and develop an EM algorithm that stratifies the partially censored data into homogeneous groups. The transition dynamics of all the subjects within a stratified group will be similar; thus, a predictive model built for a group will have a better accuracy. We demonstrate the validation and application on synthetic and real data sets. 4 - Using an Integrated System Approach in Finding the Optimal Simulation Model Based on Healthcare Case Study Tannaz Khaleghi, Wayne State University, Detroit, MI, 48202, United States, fj6710@wayne.edu, Alper E. Murat Nowadays predicting the system architecture and its corresponding core attributes at a specific time slice is valued the most. Simulation experts usually simulate the system based on the data they receive and observing some processes. However, in many cases one may not be able to gather prior information about the system design. In this study the attempt is to propose 3-stage analysis namely, process mining, dynamic Bayesian network and discrete event simulation, to discover the best system design model which is not only exclusively grantees infinitesimal deviation from the real model, but also reduces the time complexity of finding the optimal model significantly for large systems. 5 - Data Analytics & Modeling in Medical Imaging Analysis & Decision Making Shouyi Wang, Assistant Professor, University of Texas-Arlington, 500 West First Street, Woolf Hall 420, Arlington, TX, 76019, United States, shouyiw@uta.edu, Feng Liu, Kin Ming Puk, Rahilsadat Hosseini The advances in computing and advanced data analytics techniques have led to powerful and promising data-driven tools to solve complex healthcare and medical decision making problems based on increasingly large amounts of data. This session provides a collection of talks in the field to present several most recent research studies that develop data-driven approaches and machine learning techniques for comprehensive decision modeling in medical imaging analysis and decision making. 6 - Application of Conditional Binary Decision Tree to Develop Improved Hospital Length of Stay Prediction Samuel Davis, PhD Candidate, Northeastern University, Boston, MA, United States, davis.sam@husky.neu.edu Managing the variable demand and relatively fixed supply of hospital bed and nurse resources is critical for maintaining safe and efficient operations. Level- loading daily patient census requires a firm understanding of patient arrival and length of stay (LOS) patterns. This work shows that traditional LOS prediction methods are less accurate at the individual and aggregate level than the proposed method, and does so by using a novel conditional binary decision tree. Predictions are produced and validated using data from a large Boston metro area hospital. Aggregate demand predictions are created from the individual patient predictions and are shown to be accurate and rigorous. 371F 7:30 - 8:15 Syncopation /8:15 - 9:30 Mathworks Invited: Vendor Tutorial Invited Session 1 - DPL Portfolio and DPMX(tm): A Decision Analysis based System for Better Portfolio Decisions Chris Dalton, Syncopation Software, Inc, Bangor, ME, United States, cdalton@syncopation.com This demonstration will show how the DPMX(tm) System can improve portfolio decision quality by bringing better analytics and data management to your portfolio analysis process. We’ll start with an overview of DPL Professional, a widely used modelling environment for decision analysis, risk analysis and Monte Carlo simulation. Next, we’ll cover DPL Portfolio, the platform for portfolio analysis, visualization and prioritization. Finally we’ll show DPMX, a web-based based system for managing project data and presenting portfolio results in attractive, management-ready reports. The motivating examples will be drawn from the prioritization of an R&D portfolio. TA71

2 - Data Analytics with MATLAB Mary C.Fenelon, MathWorks, 3 Apple Hill Dr., Natick, MA, 01760-2098, United States, mary.fenelon@mathworks.com Join us to see how MATLAB can help you explore data, develop predictive and prescriptive models, and integrate analytics into enterprise applications. We’ll illustrate this through a load forecasting and unit commitment demo on Big Data following this workflow. Merge historical demand and weather data into a time series. Explore machine learning models interactively to find the best model to predict demand. Prescribe how to satisfy demand by formulating and a solving mixed-integer linear program. Deploy an application to the web or enterprise systems to empower business analysts. 372A Network Dynamics and Inference Sponsored: Optimization, Network Optimization Sponsored Session Chair: Chris Quinn, Purdue University, Purdue, West Lafayette, IN, United States, cjquinn@purdue.edu 1 - Identification of Peer Effects in Networked Panel Data Daniel Rock, Massachusetts Institute of Technology, Sloan School After product adoption, consumers make decisions about product use. These choices can be influenced by peer decisions in networks, but identifying causal peer influence effects is challenging. Extending the work of Bramoullé et al. (2009), we apply proofs of peer effect identification in networks under a set of exogeneity assumptions to the panel data case. With engagement data from simulation and two mobile applications, we estimate usage peer effects, comparing the performance of a variety of regression models under simulated homophily conditions. Though violation of exogeneity assumptions can bias regression estimates, there are contexts where such IV analysis can be usefully applied. 2 - Coevolutionary Latent Feature Processes for Continuous-time User-item Interactions Le Song, Georgia Institute of Technology, Atlanta, GA, United States, lsong@cc.gatech.edu Matching users to the right items at the right time is a fundamental task in recommendation systems. Traditional models based on static latent features or discretizing time into epochs become ineffective in capturing the fine-grained and complex temporal dynamics in the time-varying user-item interaction networks. We propose both parametric and nonparametric coevolutionary latent feature process models which accurately captures the coevolving nature of users’ and items’ features. Extensive experiments on diverse real-world datasets demonstrate significant improvements in user behavior prediction compared to state-of-the- arts. 3 - A Relational Event Approach to Modeling Continuous Network Change Aaron Schecter, University of Georgia, Athens, GA, United States, aschec@gmail.com, Noshir Contractor The increasing availability of continuous event data, such as emails or clickstreams, has made interaction processes significantly more visible. These interactions are driven by short-term and long-term patterns of directed events. To analyze this data with traditional network approaches, event data would have to be transformed into static links. However, this process would reduce the sequential and temporal information available. We present a method which directly utilizes this data, the relational event model, to infer the effect of event patterns on network change. 4 - Network Model Selection and Tuning by Cross-validation Tianxi Li, University of Michigan, Ann Arbot, MI, United States, tianxili@umich.edu Though many models and methods are now available for network analysis, model selection and tuning remain challenging. Cross-validation, a general tool for these tasks in many settings, is not directly applicable to networks since network splitting destroys some of the network structure. We propose a network cross- validation strategy which avoids losing information and is applicable to a wide range of network problems. We provide a theoretical justification for our method. Numerical results on both simulated and real networks show that our approach performs well for a number of model selection and parameter tuning tasks. TA72 of Management, 100 Main Street, Cambridge, MA, 02142, United States, drock@mit.edu, Sinan Aral, Sean J. Taylor

296