INFORMS 2021 Program Book

INFORMS Anaheim 2021

TE06

3 - Inviting Celebrities to Live Commerce: More Sales or More Returns? Qi Yan, University of California, Irvine, Irvine, CA, United States, L. Robin Keller Live commerce, a new popular way of shopping, is a combination of e-commerce and live streaming. Professional sellers introduce and make sales of products during live streaming. Sometimes, they invite celebrities to join the live video stream because of various reasons. Fans of the celebrities will follow the stream and make purchases in real time to validate their popularity. The large flow to the streaming channel helps the streamer become more acknowledged. However, more returns may occur after the event because of the cancellations of transactions from the fans. The net effect of inviting celebrities joining live commerce and how fit they are to the products need to be examined. TE08 CC Room 303C In Person: Applications of Markov Decision Processes General Session Chair: Daniel F Silva, Auburn University, Auburn, AL, 36849, United States 1 - Analysis of Overdiagnosis in Cervical Cancer Screening Using an Incidence-based Personalized POMDP Raha Akhavan, Sabanci University, Istanbul, Turkey, Malek Ebadi Population level screenings with a fixed frequency assumes homogeneity of the patients in different risk factors. This assumption causes unnecessary screenings and follow ups and impose considerable burden on the patients and healthcare systems. In this study, we consider different cohorts of patients with different prevalence and age-specific incidence rate of the infection, and aim to study the overdiagnosis of cervical infections and low grade lesions using a POMDP model tailored to incorporate incidence rate for each cohort. Our primary analysis confirms the presence of overdiagnosis in low and medium risk cohorts. 2 - Revisiting Linear Programming to Solve Markov Decision Processes under the Long-run Average Reward Criterion Daniel F. Silva, Auburn University, Auburn, AL, 36849, United States We compare the computational performance of Linear Programming and the Policy Iteration Algorithm for finding optimal solutions to discrete time, infinite- horizon, unichain Markov decision processes under the long-run average reward criterion. We compare the computational performance of the linear programming method and the policy iteration algorithm over test instances with varying sizes of state space, action space, as well as different sparsity and structure of the transition probability matrices. The results of our experiments show that linear programming methods are faster than the policy iteration algorithm for problems with relatively small action spaces and large state spaces, while the policy iteration algorithm is faster for problems with small state spaces and large action spaces. TE09 CC Room 303D In Person: Recent Advances in Load Balancing General Session Chair: Martin Zubeldia, Eindhoven University of Technology, Eindhoven, 5611 SJ, Netherlands 1 - Load Balancing System in the Many-server Heavy-traffic Regime Daniela Hurtado Lange, Georgia Institute of Technology, GaTech, Atlanta, GA, United States We study the load balancing system in the many-server heavy-traffic regime. We consider a system with N servers, and we parametrize the arrival rate so that the arrival rate per server is N-a, where a>0 is a parameter that represents how fast the load grows with respect to the number of servers. In this talk, we show conditions on a so that the average queue length scaled by N1-a behaves similarly to the classical heavy-traffic regime. We provide two proofs to our result: one based on Transform methods and one based on Stein’s method. In the second proof, we also compute the rate of convergence in Wasserstein’s distance. We additionally compute the rate of convergence in expected value. All of our proofs are powered by state space collapse.

TE06 CC Room 303A In Person: Community-Based Operations Research General Session Chair: EunSu Lee, New Jersey City University, Jersey City, NJ, 07304- 4048, United States Co-Chair: Michael P Johnson, University of Massachusetts Boston, Boston, MA, 02125-3393, United States 1 - Procurement Policies for Emergency Relief Operations Mahyar Eftekhar, Arizona State University, Tempe, AZ, 85287- 4706, United States, Scott Webster The aftermath of rapid-onset disaster is a chaotic period when emergency responders’ goal is to distribute critical items at the fastest possible time. This study proposes a few policies to minimize supply—demand mismatch in presence of multiple sources of uncertainty, and demonstrates the value of emergency funds. 2 - Designing a Community-engaged Learning Using Public OR EunSu Lee, Ph.D., New Jersey City University, Jersey City, NJ, 07311, United States This presentation discusses the community-engage learning (CEL) using the public OR. The case studies will be presented and introduce a sample syllabus and student project. The key takeaways include how to design CEL utilizing the public OR, lessons learned from the cases, and things to consider. The audience will be able to actively participate in the discussion during the presentation. 3 - Robust Multi-stakeholder Preference Elicitation and Aggregation for Treatment Prioritization During the Covid-19 Pandemic Caroline Johnston, University of Southern California, Los Angeles, CA, 90007, United States, Simon Blessenohl, Phebe Vayanos During the COVID-19 pandemic, triage committees must make ethically difficult decisions that are complicated by diverse stakeholder interests. We propose an automated approach to support group decisions by recommending a policy to the group a compromise between potentially conflicting individual preferences. To identify a policy to best aggregate individual preferences, our system elicits preferences by asking a moderate number of strategically selected queries, each taking the form of a pairwise comparison posed to a specific stakeholder. We propose a novel multi-stage robust optimization formulation of this problem. Formulating this as an MILP, we evaluate our approach on the issue of recommending policies for allocating ICU beds to patients with COVID-19. We show that our method recommends a policy with higher utility than various methods from the literature. TE07 CC Room 303B In Person: Novel Behavioral Models in Social Networks General Session Chair: Tauhid Zaman, Yale University, Boston, MA, 02114, United States 1 - Social Media Sentiment and Cryptocurrencies Khizar Qureshi, MIT, San Francisco, CA, 94104, United States We conduct a study of social media activity surrounding cryptocurrencies. We collect tweets from Twitter for multiple cryptocurrencies. We also construct measures to quantify the sentiment of the tweets using transformer neural networks. We model social media interactions surrounding tweets of the coin and then fit a Poisson Regression to this data and use the estimated model parameters to construct features that quantify the virality of the coin and its long-term potential for growth. Finally, we attempt to predict which coins have massive future price movements using these virality features. 2 - The Impact of Bots in the (First) Impeachment of Donald Trump Michael J. Rossetti, Adjunct Professor, Georgetown University, Washington, DC, United States, Tauhid Zaman We study manipulation of the social media discussion surrounding the first impeachment of U.S. President Donald Trump by automated accounts, known as bots. Our dataset includes 50 million posts from 2.7 million Twitter users, covering a 60 day period from impeachment to acquittal. We identify 24,000 bots using an algorithm based on the Ising model from statistical physics. Analysis shows the bots are 100 times more active than normal users, and their follower network structure is polarized along political lines. Language analysis shows pro-Trump bots using terms related to the Qanon conspiracy theory. After quantifying bot impact using a network centrality measure we developed known as generalized harmonic influence centrality, we find that although pro-Trump bots are more numerous and active than anti-Trump bots, the anti-Trump bots have a larger daily impact.

140