INFORMS 2021 Program Book

INFORMS Anaheim 2021

TE14

2 - Job Dispatching Policies for Queueing Systems with Unknown Service Rates Weina Wang, Carnegie Mellon University, Pittsburgh, PA, United States In multi-server queueing systems where there is no central queue holding all incoming jobs, job dispatching policies are used to assign incoming jobs to the queue at one of the servers. Classic job dispatching policies such as join-the- shortest-queue and shortest expected delay assume that the service rates and queue lengths of the servers are known to the dispatcher. In this work, we tackle the problem of job dispatching without the knowledge of service rates and queue lengths, where the dispatcher can only obtain noisy estimates of the service rates by observing job departures. This problem presents a novel exploration- exploitation trade-off between sending jobs to all the servers to estimate their service rates, and exploiting the currently known fastest servers to minimize the expected queueing delay. We propose a bandit-based exploration policy that learns the service rates from observed job departures. Unlike the standard multi- armed bandit problem where only one out of a finite set of actions is optimal, here the optimal policy requires identifying the optimal fraction of incoming jobs to be sent to each server. We present a regret analysis and simulations to demonstrate the effectiveness. TE10 CC Room 304B In Person: AI for Cybersecurity General Session Chair: Tung Cu, Northeastern Illinois University, Chicago, IL, 60625- 4625, United States 1 - How we Browse: Measurement and Analysis of Digital Behavior Yuliia Lut, Columbia University, New York, NY, United States, Michael Wang, Elissa Redmiles, Rachel Cummings In this work, we design and conduct a user study to collect browsing data ($n=31$) continuously for 14 days and self-reported browsing patterns. We combine self-reports and observational data to provide an up-to-date measurement study of online browsing behavior. We use these data to empirically address the following questions: (1) Do structural patterns of browsing differ across demographic groups and types of web use?, (2) Do people have correct perceptions of their behavior online?, and (3) Do people change their browsing behavior if they are aware of being observed? In response to these questions, we find significant differences in level of activity based on user age, but not based on race or gender. We also find that users have significantly different behavior on Security Concerns websites, which may enable new behavioral methods for detection of security concerns online. 2 - AI-Driven Cybersecurity: An Assessment of Cybersecurity Defense Systems Tung Cu, Northeastern Illinois University, Chicago, IL, 60625- 4625, United States Cybersecurity is simply classified into four popular categories including Data Security, Information Security, Network Security, and Internet/IoT Security. To solve these cybersecurity problems, people usually use popular AI techniques involving machine learning and deep learning methods, the concept of natural language processing, knowledge representation and reasoning, as well as the concept of knowledge or rule-based expert systems modeling. Based on these AI methods, in this paper, I present a comprehensive assessment on how these AI Cybersecurity methods can play an important role in cybersecurity defense systems. In conclusion, I also highlight several research directions within the scope of our study, which can help researchers do future research in the area. TE13 CC Room 201A In Person: Parameter Optimization in Modeling and Complex Systems General Session Chair: Mayank Kejriwa, University of Southern California, CA, University of Southern California, CA 1 - Identifying Climate Action Language in News to Predict Firm Environmental Performance Nima Safaei, University of Iowa, Iowa City, IA, United States, Gautam Pant Firms are increasingly engaged in meaningful actions to mitigate climate change effects and reduce greenhouse gas (GHG) emissions. Information on a firm’s actions to address environmental challenges, i.e., its environmental performance is critical to its stakeholders (e.g., investors, consumers, government, community). However, the currently available data about firms’ climate actions[1] is sparse and unreliable. Hence, there is a gap between the importance of firms’ climate action data and the availability of such data to their various

stakeholders. To close this gap, we develop a textual model using machine learning to derive a climate action vocabulary. The vocabulary is subsequently used for measuring the extent to which news stories about a firm are related to climate action. Specifically, we propose several text-based metrics that capture the intensity of media attention on climate action related to a given firm. We validate the effectiveness of our methodology for deriving the climate action vocabulary and the proposed text-based metrics by demonstrating that they effectively predict the future environmental performance of firms. 2 - Building Domain-specific Knowledge Graphs from Text Mayank Kejriwal, University of Southern California Domain-specific knowledge graph (KG) construction is an active and interdisciplinary research area that has recently witnessed impressive advances due to machine learning techniques like deep neural networks and word embeddings. Today, there is a vast amount of text data available, including text extracted from webpages, natural language documents, social media ‘documents’ such as tweets and memes, and even free-text descriptions in spreadsheets and CSV files. Constructing KGs over raw text data is a challenging problem that requires techniques from multiple fields, most notably Natural Language Processing. This talk will provide a high-level overview of KG construction from text, and promising directions for future research. 3 - Identifying User Needs from Online Reviews with Albert and a Cluster Based Algorithm Yi Han, Northeastern University, Malden, MA, United States Sentiment analysis has been widely applied in product design processes. Various rule-based and deep learning-based methods have been proposed in the literature to collect, analyze, and predict potential user needs from online reviews. Pre- trained language models like BERT are proven effective in different NLP tasks such as sentiment classification with over 90% accuracy. This paper utilizes ALBERT, a light improved version of BERT, with integrated domain knowledge and syntax rules to identify user needs from unstructured reviews. The ALBERT model achieves a 91.3% F1 score on a large review dataset. A cluster-based algorithm has been designed for the post-training of the model. 4 - Quantifying Risks of Food Insecurity by Analyzing the News Ananth Balashankar, Ph.D Candidate, New York University, New York, NY, United States, Samuel Fraiberger, Lakshminarayanan Subramanian Existing food security early warning systems rely on sparse data, usually available with a considerable lag, impeding the efficacy of humanitarian efforts. In this work, we propose a novel framework to extract early indicators reported in news articles, which are both causally grounded and highly predictive of food insecurity events. Using predictive causal early indicators from news, we reduce the error in forecasting the average Famine Early Warning System rating in 20 fragile states by 37% as compared to using on-the-field data alone. Analyzing news from 30 years, we also show that early detection of certain social and political causes of famine, improves the AUC-PR of crisis events by 20%. TE14 CC Room 201B In Person: Machine Learning for Spatio-temporal Data Modeling and Analysis General Session Chair: Shixiang Zhu, Georgia Institute of Technology, Atlanta, GA, 30318-2990, United States 1 - Equivariant Neural Networks for Learning Spatiotemporal Dynamics Robin Walters Applications such as climate science and transportationrequire learning complex dynamics from large-scale spatiotemporal data.Existing machine learning frameworks are still insufficient to learnspatiotemporal dynamics as they often fail to exploit the underlyingphysics principles. Representation theory can be used to describe andexploit the symmetry of the dynamical system. We will show how to designneural networks that are equivariant to various symmetries for learningspatiotemporal dynamics. Our methods demonstrate significantimprovement in prediction accuracy, generalization, and sampleefficiency in forecasting turbulent flows and predicting real- worldtrajectories. This is joint work with Rose Yu, Rui Wang. and Jinxi Li. 2 - An Adaptive Sampling Strategy for Online Monitoring and Diagnosis of High-dimensional Streaming Data This research proposes an adaptive sampling strategy for online monitoring and diagnosis of high-dimensional streaming data. It integrates two novel ideas: (i) the recursive projection of the high-dimensional streaming data onto a low- dimensional subspace to capture the spatio-temporal structure of the data while performing missing data imputation, and (ii) the development of an adaptive sampling scheme, balancing exploration and exploitation. Kamran Paynabar, ISyE Georgia Tech, Georgia Tech, H. Milton Stewart School of Isye, Atlanta, GA, 30332-0205, United States

141