INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME41

2 - Can Autonomous Vehicles Solve the Commuter Parking Problem? Neda Mirzaeian, Carnegie Mellon University, 5738 Wilkins Ave Apt 1a, Pittsburgh, PA, 15217-1249, United States, Soo-Haeng Cho, Sean Z. Qian We investigate the effect of autonomous vehicles (AVs) on the morning commute. We characterize a user equilibrium for commuters by developing a continuous- time traffic model that takes into account parking fees and traffic congestion as two key economic deterrents to driving. In addition, we consider the case of a social planner who aims to minimize the total system cost by controlling the commuters’ decisions. We illustrate our results using data from Pittsburgh, and show that AVs reduce the need for downtown parking. We also show that, in the short run, adjusting downtown parking fees and imposing tolls on downtown congestion can reduce the total system cost significantly (e.g., a 51% reduction for Pittsburgh), and that adjusting the road and parking capacities as a long-term plan may reduce the total cost even further (e.g., an additional 70% reduction for Pittsburgh). ME37 CC Room 210C In Person: Managerial Issues in the Platform Economy General Session Chair: Koushiki Sarkar, Northwestern University, Evanston, IL, 60201- 4434, United States 1 - Why Harry Wouldn’t Meet Sally: An Empirical Analysis of Gender Disparity in Online Learning Forums Zhihan (Helen) Wang, Ross School of Business, University of Michigan, Ann Arbor, MI, United States, Jun Li, Andrew Wu On most Massive Open Online Courses (MOOC) platforms, the discussion forum is the primary source of interaction among learners. Using a large-scale data set of forum posts on a world-leading MOOC platform, we demonstrate the existence of gender bias in discussion behavior among learners, and examine the sources of this bias and its potential impact on engagement and learning outcomes. 2 - Strategic Considerations in the Presence of Social Influencers Gad Allon, University of Pennsylvania, Philadelphia, PA, 19104- 3615, United States, Koushiki Sarkar, Achal Bassamboo The use of social information for decision-making is of great practical importance in various settings, amongst which online shopping holds a special place. Customers frequently base their purchase decisions on reviews of prominent users, typically known as influencers. We address the following questions: what features promote a particular user to the level of an influencer? How should firms incentivize influential reviewers to structure their reviews? When customers have a limited search budget, we show that they do not necessarily follow the most precise or most open influencer. Further, the firm does not always benefit from disclosing accurate quality information. The firm may profit significantly by appropriately controlling the social information displayed to the agents, either by targeting influencers to post more or providing better quality information ME39 CC Room 211A In Person: New Directions in Operations Management General Session Chair: Maria R. Ibanez, Kellogg School of Management at Northwestern University, Evanston, IL, 60208-0898, United States 1 - The Power of Analytics in Epidemiology for the COVID-19; Prediction, Prevalence and Vaccine Allocation David A. Nze-Ndong, Massachusetts Institute of Technology, Cambridge, MA, United States, Mohammed Amine Bennouna, Georgia Perakis, Omar Skali Lami, Ioannis Spantidakis, Leann Thayaparan, Asterios Tsiourvas Mitigating the COVID19 pandemic poses many challenges. Those include predicting new cases and deaths, understanding true prevalence, and allocating vaccines. We present a novel predictive ML-based aggregation method (MIT- Cassandra) also used by the CDC that is consistently among the top 10 models in terms of accuracy. We then predict the true prevalence of COVID19 and incorporate it into an optimization model for fair vaccine allocation. We obtain interesting insights on how prevalence affects the vaccine distribution for a heterogeneous population. Our work has been part of a collaboration with MIT’s Quest for Intelligence and as part of CDC’s model ensemble.

2 - Supply Chain Characters as Predictors of Cyber Risk: A Machine-learning Assessment Retsef Levi, MIT, Sloan School of Man., Cambridge, MA, 02142- 1320, United States The presentation provides the first empirical evidence that certain supply-chain attributes are significant predictors of cyber risk for enterprises, in addition to their internal characteristics and level of cybersecurity management. It leverages outside-in cyber risk scores that represent quality of cyber security management, and augment these with supply chain features that are inspired by network science research, to develop a more comprehensive risk assessment. The main result is to develop a model that shows that supply chain network features add significant detection power relative to merely internal enterprise attributes in predicting risk of cyber data breach incidents. Additionally, the model highlights several cybersecurity risk insights related to third party data breach mechanisms that have seen significant increase over the last several years. 3 - An Empirical Study of Time Allotment and Delays in E-commerce Delivery Natalie Epstein, Harvard Business School, Cambridge, MA, 02138, United States, Maya Balakrishnan, MoonSoo Choi Time allotment comes with an inherent tradeoff between delays and duration, so managers need to carefully evaluate such tradeoff. We explore the relationship between time allotment and delivery outcomes in an e-commerce delivery context and further seek to identify relevant features for predicting order delays and study how real-time information of the delivery process can improve prediction accuracy. We use the JD.com transaction dataset provided by Shen et al. (2020) and find that (i) increasing the allotted time for an order increases the duration of its delivery process, (ii) more allotted time reduces the likelihood of an order being late or delayed, (iii) our delay prediction models exhibit accuracy levels higher than no-information rate, and (iv) adding information from early parts of the delivery significantly increases accuracy when predicting future delays. ME40 CC Room 211B In Person: Machine Learning in Finance General Session Chair: Renyuan Xu, University of Southern California, Los Angeles, Renyuan Xu, University of Southern California, Los Angeles, CA, 94720, United States, Alain-Sam Cohen, Rama Cont, Alain Rossier Residual networks (ResNets) have displayed impressive results in pattern recognition and have garnered considerable theoretical interest due to a perceived link with neural ODEs. This link relies on the convergence of network weights to a smooth function as the number of layers increases. We investigate the properties of weights trained by SGD and their scaling with network depth through detailed experiments. We observe the existence of scaling regimes markedly different from those assumed in neural ODE literature. Depending on certain features of the network architecture, we prove the existence of an alternative ODE limit, an SDE limit, or neither of these. These findings cast doubts on the validity of the neural ODE model as an adequate asymptotic description of deep ResNets and point to an alternative class of differential equations as a better description of the limit. ME41 CC Room 212A In Person: ARPA-E Grid Optimization Competition 2 General Session Chair: Constance Crozier, University of Colorado, Boulder, CO, United States 1 - Algorithmic Development for Solving Large-Scale Security Constrained AC-OPF with Switching Andy Sun, Georgia Tech, Atlanta, GA, United States 2 - On Solving Large-Scale Mixed-Integer Nonlinear Programs Hassan Hijazi The talk will cover recent work on solving large-scale Mixed-Integer NonLinear Programs (MINLPs). We will discuss ideas such as automatic projection of auxiliary variables, building convex outer relaxations as well as convex inner restrictions, and showcasing how they both play an important role in providing feasible solutions and optimality guarantees for challenging MINLPs. Numerical results on the ARPA-e Grid Optimization Competition Challenge 2 will be presented. CA, 94720, United States Co-Chair: Zhengyuan Zhou 1 - Scaling Properties of Deep Residual Networks

97