INFORMS 2021 Program Book

INFORMS Anaheim 2021

TD15

2 - Optimizing Pandemic Policy Decisions Accounting for Economic And Epidemiologic Impact Leili Soltanisehat, University of Oklahoma, Norman, OK, 73019- 1027, United States While different COVID-19 controlling strategies (e.g., lockdown, school, and business closures) helped with decreasing the number of infections, they have had an adverse economic impact. Conversely, loosening such strategies led to new waves of cases. Therefore, the optimal timing and scale of closure and reopening strategies are required to balance pandemic growth with economic impacts. This work proposes a novel mathematical model for mitigating the economic and epidemiologic impact of a pandemic by combining SIRD and MNFP models into a linear mixed-integer program to explore state- and industry-level strategies. 3 - COVID-19 ResNet: Residual Neural Network for COVID-19 Classification with Bayesian Data Augmentation Javier Sebastian Balseca Zurita, Universidad San Francisco de Quito, Quito, Ecuador, Martin Alejandro Cruz Patino In this work a Residual Convolutional Neural Network (ResNet) for COVID-19 medical image (CXR) classification with a personalized data augmentation strategy is presented. The ResNet is a very deep network that progressively learns high level and complex feature representations from the CXR images. To overcome the data scarcity of covid-19 images, a data augmentation approach was implemented and hyperparameters were optimized using a Bayesian optimization approach. Experimental results show the proposed method obtains a competitive classification accuracy. 4 - Multi-criteria Course Mode Selection and Classroom Assignment under Sudden Space Scarcity Mehran Navabi, ISyE Georgia Tech, Atlanta, GA, United States, Mohamed El Tonbari, Boland Natashia, Dima Nazzal, Lauren N. Steimle Social distancing dramatically reduces the effective capacity of classrooms. During the COVID-19 pandemic, this presented a unique problem to campus planners: (1) Assigning a mode to each offered class as either remote, residential (in- person) or hybrid (2) Reassigning classrooms under reduced capacities to the non-remote classes. We solve a flexible integer program and use hierarchical optimization to handle the trade off between various administrative priorities. We generate optimal classroom assignments for all classes at the Georgia Institute of Technology, and quantify the impact of our results, particularly on in-person contact hours and mode preference satisfaction. 5 - Impact Assessment of Full and Partial Stay-at-home Orders, Face Mask Usage, and Contact Tracing: An Agent-based Simulation Study Of Covid-19 for an Urban Region Shalome Hanisha Anand Tatapudi, University of South Florida, Tampa, FL, United States, Rachita Das, Tapas K. Das Social intervention strategies to mitigate COVID-19 are examined using an agent- based simulation model. The simulation model mimics daily social mixing behavior of the susceptible and infected and data representing demographics of the region, virus epidemiology, and social interventions shapes model behavior. Results show that early implementation of complete stay-at-home order is effective in flattening the infection growth curve in a short period of time. Universal use of face masks reduced infected by 20%. A further reduction of 66% was achieved by adding contact tracing with a target of identifying 50% of the asymptomatic and pre-symptomatic. TD15 CC Room 201C In Person: Data Analytics in Semiconductor Manufacturing General Session Chair: Kim Dohyun, Korea, Republic of 1 - A Framework for Process Parameter Optimization Based on Deep Generative Model YoungGeun Ahn, Myongji University, Yongin-si, Korea, Republic of, Dohyun (Norman) Kim, Minyoung Park There are many process parameters that affect the process yield in the semiconductor process, and it is very important to find an optimized value of parameters to maximize the process yield. However, complex nonlinear correlations between parameters make their optimization difficult. Therefore, in this study, we propose a framework for optimizing the process parameters that maximize the process yield using deep generative model.

2 - Deep Learning-based Clustering Algorithm Considering Outliers Somi Ha, Myong-ji Univ., Gyeonggi-do, Korea, Republic of, Sungwoo Kim, Dohyun (Norman) Kim Most clustering algorithms perform clustering without considering outliers. However, when performing clustering, outliers often degrade performance. Therefore, in this study, a clustering algorithm that simultaneously performs clustering and outlier detection has been proposed. The proposed deep learning- based algorithm identifies outliers using the predicted class distribution and performs clustering based on only normal data. 3 - Analysis of Tabular Data Based on Graph Neural Network Seungyeon Lee, Myongji University, Yongin-si, Korea, Republic of, Minyoung Park, Dohyun (Norman) Kim In many applications in the industry, tabular data are the most commonly used data type. Machine learning methods for dealing with tabular data are classified into two categories: similarity-based approach and feature-based approach. Feature-based models are easy to understand and intuitive to use and deploy but generally cannot utilize the relationships between observations. Similarity-based models are most suited for exploiting the relationships among observations, but their availability is usually limited. In order to take advantage of both aspects of tabular data, we propose an algorithm to combine feature-based and similarity- based approaches using graph neural network. Experimental results show that the proposed method provides more precise results for classification tasks, implying that it may improve the generalization capability. 4 - Competing Streaming Platforms: The Impact of Exclusive Content Emily A. Meigs, MIT, Cambridge, MA, 02139-4204, United States We develop a model to study the joint problem of designing the subscription fee and amount of original content a streaming platform should generate. In our model, the users are heterogeneous in their usage rate and depending on the content on the platforms and their subscription fee decide whether they want to subscribe to a platform or not. The two competing platforms, each choose the subscription fee that they want to offer for (unlimited) access to their content and the investment level they put into their content. The investment in their own content is costly, but higher quality content can potentially be offered at a higher subscription fee to the users. We fully characterize the equilibrium in both the monopolistic and competitive settings. We characterize under what conditions the platforms separate the two types of customers. 5 - A New Cluster Validity Index for Non-hierarchical Clustering Analysis Youngseon Jeong, Chonnam National University, Industr, Gwangju, 61186, Korea, Republic of, Changwan Ko This paper proposes new cluster validity indexes (CVIs) in feature space for non- hierarchical clustering analysis in which the proposed CVIs transform arbitrary shape of clusters into elliptical or circular shaped clusters by using kernel functions of Support Vector Data Description (SVDD). The experimental results show that the proposed CVIs have a good performance to estimate the optimum number of clusters for lower-dimensional and unique characteristic dataset. TD16 CC Room 201D In Person: Optimization Methods for Learning from Data General Session Chair: Paul Brooks, Virginia Commonwealth University, Richmond, VA, 23284-4000, United States Co-Chair: Jose H. Dula, University of Alabama, Tuscaloosa, AL, 35487, United States 1 - Nearest Convex Hull Classification with Linear Programming The multi-class classification problem aims at assigning a test point to one of several classes that partition a data set. Nearest Convex Hull Classification uses the point’s distance to the convex hulls of the class’s data for this assignment. This presents a special challenge when the test point is interior to two or more hulls. We propose a lazy supervised machine learning method based on linear programming that locates internal and external test points and approximates distances. Advantages include that the same formulation is used for interior or exterior points, necessary and sufficient conditions for classification, the absence of user-defined parameters, and excellent scalability. Tests on health care data show the method performs well. Jose H. Dula, University of Alabama, School of Business,, Tuscaloosa, AL, 35487, United States, Anatoly Nemirko

130