Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TA62

strongly convex, convex, and nonconvex cases. We also compare it with some existing stochastic algorithms. 2 - Asynchronous Decentralized Accelerated Stochastic Gradient Descent Yi Zhou, ISyE Georgia Tech, 755 Ferst Drive, NW, Atlanta, GA, 30332, United States, Guanghui Lan In this work, we introduce an asynchronous decentralized accelerated stochastic gradient descent type of method for solving stochastic convex optimization problems defined over multiagent networks. Considering that communication and synchronization are the major bottlenecks in decentralized optimization, our main goal in this talk is to present algorithmic frameworks which can significantly reduce communication and synchronization costs. We established ${\cal O}\{1/\epsilon\}$ (resp. ${\cal O}\{1/\sqrt \epsilon\}$) rate of convergence for our proposed method when the problem is general convex (resp. strongly convex). 3 - L0-regularized Sparsity for Probabilistic Mixture Models Dzung Phan, IBM Research, 1101 Kitchawan Road, P.O. Box 218, Yorktown Heights, NY, 10598, United States, Tsuyoshi Ide This talk revisits the task of learning probabilistic mixture models. Our major goal is to sparsely learn the mixture weights to automatically determine the right number of clusters. The key idea is to use a novel Bernoulli prior on the mixture weights in a Bayesian learning framework, and formalize the task of determining the mixture weights as an L0-regularized optimization problem. By leveraging a specific mathematical structure, we derive a quadratic time algorithm for efficiently solving the non-convex L0-based problem. In experiments, we evaluate the performance of our proposed approach over existing sparse methods in terms of accuracy and scalability on synthetic and real data sets. 4 - Scaling Up Optimization Using Non-convexity, Provably Anastasios Kyrillidis, 1616 Guadalupe Street, Austin, TX, 78701, United States In this talk, I will focus on the problem of low rank matrix inference in large-scale settings. Such problems appear in fundamental applications such as structured inference, recommendation systems and multi-label classification problems. I will introduce a novel theoretical framework for analyzing the performance of non- convex first-order methods, often used as heuristics in practice. These methods lead to computational gains over classic convex approaches, but their analysis is unknown for most problems. This talk will provide precise theoretical guarantees, answering the long-standing question “why such non-convex techniques behave well in practice”? for a wide class of problems. Chair: Durai Sundaramoorthi, Washington University in Saint Louis, Washington University in Saint Louis, Saint Louis, MO, 63131, United States Co-Chair: Sundar Victor, Aetna, New York, NY, United States 1 - Study of Mobility in Private Homes Yan Wang, University of South Florida, Industrial and Management Systems Engineering, 4202 E. Fowler Ave, ENB118, Tampa, FL, 33620, United States, Ali Yalcin, Carla VandeWeerd Human mobility is shown to be fundamentally regular and potentially predictable based on the study of large-scale geo-location data available through cellular networks and mobile devices. In this work, we study the mobility of residents in private homes, as opposed to outdoor environments. Indoor mobility prediction has far-reaching applications in smart homes and ambient assisted living environments. Using data from ambient sensors, we represent the occupant’s movement trajectory as a sequence of symbols and calculate its entropy and an upper bound of its predictability. Our results show that indoor mobility is also predictable and can be quantified. 2 - Deep Neural Networks and Insurance Cross-selling Xiaoguang Tian, University of North Texas, 1155 Union Circle, Denton, TX, 76203-5017, United States In this study, we propose applying deep neural networks and ensemble approach to identify the customer who will purchase a new caravan insurance policy and improve the model performance through tuning the parameters. The results show that the approach is effective and outperforms the baseline model on multiple measurements, such as accuracy, sensitivity, and ROC. n TA64 West Bldg 104A Joint Session DM/Practice Curated: Data Science for Novel Applications Sponsored: Data Mining Sponsored Session

n TA62 West Bldg 103A Mining Unstructured Data Sponsored: Data Mining Sponsored Session

Chair: Erhun Kundakcioglu, Ozyegin University, Istanbul, Turkeyr Co-Chair: Mustafa Gokce Baydogan, Bogazici University, Istanbul, 34342, Turkey 1 - Nested Gaussian Process for Unstructured Data Mining Hui Yang, Pennsylvania State University, 310 Leonhard Building, Industrial and Manufacturing Engineering, University Park, PA, 16801, United States Modern healthcare systems are increasingly investing in advanced sensing and information technology, leading to data-rich environments in hospitals. However, heterogeneous types of sensing and measurement methods often result in data uncertainty and incompleteness. Missing values are not uncommon for clinical variables pertinent to a patient’s health conditions. This adversely affects data- driven decision making in the healthcare environment. This paper presents a novel nested Gaussian process (NGP) model that is tailored to represent multi- dimensional covariance structure of time, variable and patient for high-dimensional unstructured data mining in healthcare systems. 2 - Understanding Student Enrollment through Unstructured Test Response Data Analysis Petros Xanthopoulos, Stetson University, 421 Woodland Blvd, DeLand, FL, 32723, United States, Berry T. Cox Free text response is very important in education analytics. Survey responses might encapsulate information and insights that could help an institution improve its services and adjust better to the needs of future students. In this presentation we present out analysis for understanding the connection between early post campus visit student surveys and university enrollment. In our study we incorporate natural language processing and machine learning in order to uncover markers that would help us predict factors of importance for predicting student enrollment. 3 - A Sparse Graph-regularized Cox Model for Hot Spot Search on 3-D Radiation Dose Map in Cardiac Toxicity Assessment of Lung Cancer Xiaonan Liu, Arizona State University, 699 S. Mill Avenue, Tempe, AZ, 85281, United States, Mirek Fatyga, Jing Li Recent clinical trials on dose escalation in radiation therapy of lung cancer found decrease in overall survival (OS). This unexpected result has driven the research society to study potential damage of radiation to the heart, known as cardiac toxicity. Existing studies focus on finding features in dose-volume histogram (DVH) that are correlated with the toxicity or OS. However, DVH collapses the 3- D dose map into 2-D and thus losing spatial information of dose distribution. We propose a novel sparse graph-regularized cox model to find localized regions (hot spots) in the 3-D dose map of the heart to predict OS. This work enables better treatment planning for lung cancer with awareness of cardiac risk. 4 - Maximizing Area Under Receiver Operator Characteristic Curve for Multi-Instance Learning Erhun Kundakcioglu, Ozyegin University, Faculty of Engineering, Nisantepe mah. Orman sok., Istanbul, 34794, Turkey The purpose of this study is to solve the multi-instance classification problem by directly maximizing the area under ROC curve. We derive a mixed integer linear programming model that produces the best possible hyperplane-based classifier for multi-instance classification and present cross validation (CV) results for benchmark instances. n TA63 West Bldg 103B Recent Advances in Structured and Non-convex Optimization for Machine Learning Sponsored: Data Mining 1 - Inexact SARAH for Solving Stochastic Optimization Problems Lam Nguyen, IBM Thomas J. Watson Research Center, Yorktown Heights, NY, United States, Katya Scheinberg, Martin Takac We consider a general stochastic optimization problem which is often at the core of supervised learning, such as deep learning and linear classification. It is not able to apply SVRG and SARAH for this problem since these algorithms require an exact gradient information. We consider Inexact SARAH, which does not require to compute an exact gradient at each outer iteration; and analyze it in Sponsored Session Chair: Dzung Phan

266