Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MA64

2 - Effects of User-provided Photos on Hotel Review Helpfulness: An Analytical Approach with Deep Leaning Qianzhou Du, Virginia Tech, Blacksburg, VA, United States, Weiguo Fan, Zheng Xiang Online reviews have been studied in the hospitality and tourism literature. However, while user provided photos embedded in online reviews accumulate in large quantities, their informational value has not been well understood likely due to technical challenges. The goal of this study is to introduce deep learning for computer vision to understand information value of online hotel reviews. Using a dataset collected from two social media sites, we compared deep learning models with other machine learning techniques to examine the effect of user- provided photos on review helpfulness. Findings show that deep learning models are more useful in predicting review helpfulness than other models. 3 - Deep Learning-based Sentence-level User Feedback Classification in Mobile App Reviews Zhilei Qiao, Virginia Institute of Technology, 880 West Campus Drive, Blacksburg, VA, 24061, United States, Alan Gang Wang, Alan Abrahams, Weiguo Fan It is important for app developers to efficiently extract and understand user needs from online reviews. However, discovering and quantifying potential user needs from massive online reviews is a nontrivial task. In this paper, I propose a domain-oriented deep learning framework that can discover the most critical user needs such as new product features and bug reports from online reviews. I conduct a systematic evaluation to ensure the quality of discovered information. Experimental results demonstrate that the proposed framework outperforms the baseline models. This research has significant managerial implications for app developers. 4 - Understanding Meaning of Text in its Context – A Network-based Text Mining Model for Discovering Context Clues Sukhwa Hong, Virginia Institue of Technology, Pamplin College of Business, 880 West Campus Drive, Blacksburg, VA, 24061, United States, Onur Seref, Michelle MH Seref We present a network-based framework to identify and cluster phrases or phrases with context clues in classes of text data using prevalence scores of n-gram structures and their connections. Our framework extends the bag-of-words models by network-based clustering methods to create sub-graphs of connected n-grams for finding context clues. The paths in these sub-graphs represent sequences of words, which form connected phrases with richer contextual meaning. We use our method to identify variations of these phrases and apply the proposed framework to study a collection of customer reviews from TripAdvisor and Yelp. n MA64 West Bldg 104A Joint Session DM/Practice Curated: Data Science and Analytics in Healthcare II Sponsored: Data Mining Sponsored Session Chair: Durai Sundaramoorthi, Washington University in Saint Louis, 10352 Conway Road, Saint Louis, MO, 63131, United States 1 - Machine Learning Based Hypoglycemia Recognition from Driving Patterns in Individuals with Diabetes Mellitus Mathias Kraus, ETH Zuerich, Zuerich, Switzerland, Stefan Feuerriegel, Elgar Fleisch, Tobias Kowatsch, Markus Laimer, Christoph Stettler, Felix Wortmann, Thomas Z ger Hypoglycemia has consistently been shown to be associated with an increased risk of driving mishaps. As a prevention, we propose to utilize machine learning models to detect and predict hypoglycemia based on driving behavior. We compare data about the driving behavior by tracking individuals with/without hypoglycemic condition. Our preliminary evaluation using a driving simulator confirm the effectiveness of our approach. 2 - Using Machine Learning Techniques to Determine Preterm Birth Risk Factors Alireza Ebrahimvandi, Virginia Tech, Blacksburg, VA, 24060, United States Preterm birthùbirth before 37 wksùis a major health issue in the US and its predictors are poorly known. Prior studies applied machine learning techniques to identify variables that can predict preterm. We will use two machine learning techniquesùlasso regression and support vector machinesùon a much larger dataset obtained from Center for Disease Control to identify key predictors of preterm. We expect to improve sensitivity of prediction by using a larger dataset. Improving one percent sensitivity of prediction tests is valuable because every preterm birth costs about $10,000, and 1% decrease in preterm births translates to half a billion dollars saving plus preventing long life deficiencies.

3 - Combining Observational Data and Meta-analysis Results for Evaluating Impact of Behavior Changes on Disability Adjusted Life Years Ozden F. Gur Ali, Koc University, College of Administrative Sciences, Rumeli Feneri Yolu Sariyer, Istanbul, 34450, Turkey, Angi Ghanem We introduce a method to combine individual level observational data with meta- analysis results of extant research to evaluate the potential impact of a public health intervention. We show that both pieces of information are needed to get causally defensible models that reflect the local effects, adjust for important individual level covariates, and guard against confounding. We apply the method to provide point and interval estimates of the impact of changes in behaviors like physical exercise, smoking, and diet on disability adjusted life years (DALY) due to prevalence of heart disease and diabetes. n MA65 West Bldg 104B Joint Session APS/Opt-Uncert/Practice Curated: APS Session Uday Sponsored: Applied Probability Sponsored Session Chair: Uday Shanbhag, Pennsylvania State University, University Park, PA, 16803, United States 1 - Randomized Push-pull Method for Optimization on Graphs Angelia Nedich, ASU, 650 E. Tayler Mall, Goldwater Center, Rm 311, Tempe, AZ, 85281, United States We consider a randomized push-pull algorithm for distributed optimization on graphs. The method is based on a random asynchronous implementation of push- pull method recently proposed for minimizing the sum of the agents’ cost functions in a network that obeys the network connectivity structure. In order to minimize the sum of the cost functions, a new distributed gradient-based method where each node maintains two estimates, namely, an estimate of the optimal decision variable and an estimate of the gradient for the average of the agents’ objective functions. We show convergence results for the method over a directed static network. 2 - Estimating Optimality Gaps for Stochastic Optimization via Bootstrap Aggregating Henry Lam, Columbia University, 500 W. 120th St., New York, NY, 10027, United States, Huajie Qian We present approaches to estimate the optimality gaps of solutions in stochastic optimization using limited data. Our approaches are based on bootstrap aggregating that can demonstrably improve the tightness-variance tradeoff incurred in existing methods. We present our theories and show numerical effectiveness through some examples. 3 - Asymptotic Results on Two-stage Stochastic Quadratic Programming Junyi Liu, University of Southern California, 30 Virginia Avenue, Los Angeles, CA, 91107, United States, Suvrajeet Sen In this talk, we will present stochastic decomposition (SD) algorithms for two- stage stochastic quadratic programming (STQP) problem. Based on their stochastic linear programming (SLP) predecessor, these iterative schemes in SD algorithm approximate the objective function using affine/quadratic minorants and then apply a stochastic proximal mapping to obtain the next iterate. We show that under certain assumptions the proximal mapping applied in SD obey a non- expansive mapping property and study the convergence rate of SD to the optimal solutions. 4 - Smoothing and Acceleration for Stochastic Convex Optimization Afrooz Jalilzadeh, Pennsylvania State University, 310 Leonhard Building, University Park, PA, 16803, United States, Uday Shanbhag We consider a class of structured nonsmooth stochastic convex problems. By integrating smoothing, Nesterov acceleration, and increasing batch-sizes, we show that Nesterov’s rate of O(1/k) may be recovered while under strong convexity, a linear rate of convergence is attained. We comment on special cases in the convex regime when the problem is either smooth or merely deterministic.

142