INFORMS 2021 Program Book

INFORMS Anaheim 2021

WC10

2 - A Stochastic Control Approach to Quasi-stationary Distributions Pierre Nyquist, KTH Royal Institute of Technology, KTH Royal Institute of Technology, Department of M, Stockholm, 10044, Sweden, Amarjit Budhiraja, Paul Dupuis, Guo-Jhen Wu Quasi-stationary distributions (QSDs) are a core concept within applied and computational probability. For example, they are at the heart of the study of population processes, and for systems exhibiting metastability, QSDs determine important quantities such as mean exit times and exit points from metastable states. In this talk, I will introduce a new approach for studying QSDs based on ergodic stochastic control problems, in the setting of diffusions on a bounded domain. I will describe the link between QSDs and such control problems, along with how the associated Hamilton-Jacobi-Bellman equations can be used to characterise important properties of the QSD. Time permitting, I will also mention briefly how this connection can be used to construct efficient numerical schemes, and understand and explain non-uniqueness of QSDs in unbounded domains. WC10 CC Room 304B In Person: AI in New Business Models General Session Chair: Jeffrey Clement, Minneapolis, MN, 55406, United States 1 - The New and the Reliable: Novelty, Credibility, and Helpfulness in Online Reviews Dicle Yagmur Ozdemir, The University of Texas at Dallas, Dallas, TX, 75080-3021, United States, Harpreet Singh, Sumit Sarkar Online reviews and ratings are important for online platforms. To better leverage such content, platforms enable users to vote on the helpfulness of reviews. Factors found to impact the helpfulness of a review include, among others, the novelty of the content in the review and the review’s credibility characteristics (i.e., source credibility and rating credibility). We investigate the moderating impact of credibility on the effect of review novelty on helpfulness. We find that source credibility and review novelty are substitutes in terms of their contribution to review helpfulness. On the other hand, rating credibility positively moderates the effect of a review’s novelty on its helpfulness. 2 - Information Transparency and Market Efficiency in Blockchain- A unique feature of Blockchain-enabled marketplaces is that detailed account- level transaction data and the entire trading history of products are publicly available. This study examines how transparent information influences market efficiency in Blockchain-enabled marketplaces, and how good people make use of transparent information. We reveal that more information is not necessarily an indicator for higher market efficiency, which may hinge on who the traders are, especially their analytical ability to make use of the transparent information. If traders do not have the ability to utilize information effectively, market inefficiency can persist. Our study reveals that the problem of ability divide instead of digital divide prevails nowadays: digital technology has been available to all, but how to make people use technology effectively is the impending problem. 3 - Disregarding, Modifying, and Adopting: How Medical Experts Incorporate AI Recommendations Into Patient Care Decisions Jeffrey Clement, University of Minnesota, Minneapolis, MN, United States, Yuqing Ren, Shawn P. Curley AI Clinical Decision Support Systems (AI CDSS) can generate personalized recommendations to improve patient care, but it is unclear how healthcare professionals incorporate these recommendations into their care decisions. We employ mixed methods with semi-structured interviews and a pair of computer- based experiments with experienced organ transplant clinicians to examine the factors that influence trust of AI CDSS. Our results indicate that the process of incorporating AI recommendations into clinical decisions is not explained by the theories explaining trust in other recommender systems; notably, providing explanations does not seem to increase trust in the recommendations. enabled Marketplaces: Role of Traders’ Analytical Ability Hong Zhang, University of Texas at Dallas, Dallas, TX, United States

WC08 CC Room 303C In Person: Random Graphs and Learning in Applied Probability General Session Chair: Jiaming Xu, Duke University, Milano, Italy 1 - Shotgun Assembly of Erd s-Rényi Random Graphs Julia Gaudio, Northwestern University, Evanston, IL, 02139-4204, United States, Elchanan Mossel Graph shotgun assembly refers to the problem of reconstructing a graph from a collection of local neighborhoods. We consider shotgun assembly of Erdos-Renyi random graphs G(n, p_n), where p_n = n^{-\alpha} for 0 < \alpha < 1. We consider both reconstruction up to isomorphism as well as exact reconstruction (recovering the vertex labels as well as the structure). We show that given the collection of distance-1 neighborhoods, G is exactly reconstructable for 0 < \alpha < 1/3, but not reconstructable for 1/2 < \alpha < 1. Given the collection of distance-2 neighborhoods, G is exactly reconstructable for 0 < \alpha < 3/5, but not reconstructable for 3/4 < \alpha < 1. 2 - The Planted Matching Problem: Sharp Threshold and Infinite- order Phase Transition ““Ix” “Dana Yang, Duke University, Durham, NC, United States, Jian Ding, Yihong Wu, Jiaming Xu Motivated by the application of tracking moving particles from snapshots, we study the problem of reconstructing a perfect matching hidden in a randomly weighted Erd s-Rényi bipartite graph with average degree d. The edges are associated with weights independently drawn from distributions P or Q, depending on whether the edge is in the hidden matching.We establish that the information-theoretic threshold for recovering almost all the edges of the hidden matching occurs at √ dB(P,Q)=1, where B(P,Q) stands for the Bhattacharyya coefficient. Furthermore, in the special case of complete exponentially weighted graphs, we characterize the optimal reconstruction error near the sharp threshold, confirming the conjectured infinite-order phase transition in [Semerjian et al. 2020]. 3 - Detection and Recovery Thresholds for Graph Matching Sophie H. Yu, Duke University, Durham, NC, United States, Yihong Wu, Jiaming Xu This talk focuses on detection and recovery problems of matching two Erdos- Renyi random graphs.Specifically, for detection, we aim to decide whether the two observed graphs are independent, or edge-correlated under some latent node correspondence. For recovery, our goal is to recover the latent node correspondence given the two graphs are edge-correlated. In the dense graph regime, we prove that both detection and recovery exhibit an “all-or-nothing” phase transition at a sharp threshold.For sparse graphs, we identify the information-theoretic threshold within some constant factor. WC09 CC Room 303D In Person: APS session General Session Chair: Yuanlu Bai, Columbia University, New York, NY, 10027-7105, United States Co-Chair: Harsha Honnappa, Purdue University, West Lafayette, IN, 47907-2023, United States 1 - Low-rank Approximation for MDPs via Moment Coupling Amy B.Z. Zhang, Cornell University, New York, NY, 10044-1500, United States, Itai Gurvich We propose a method to approximate a Markov Decision Process that is based on state aggregation as the algorithmic infrastructure, and central-limit-theorem-type approximations as the mathematical underpinning for guarantees. The theory is grounded in recent work (Braverman et al, 2020) that relates the solution of the Bellman equation to that of a PDE where the transition matrix is reduced to its local first and second moments. We then construct an approximate “sister” chain whose local transition moments are approximately identical with those of the focal chain, coupling them through the PDE. Embedded into the framework of (soft) aggregation, moment matching motivates a disciplined mechanism to tune the aggregation and disaggregation probabilities, resulting in an efficient addition to the standard aggregation algorithm while providing optimality guarantees.

159