Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

WA69

4 - Time Allocation Strategy for Evolutionary Process in Biomanufacturing

n WA68 West Bldg 105C

Mahdi Fatihi, University of Florida, 401 Weil Hall P.O. Box 116595, Gainesville, FL, United States, Marzieh Khakifirooz, Chen-Fu Chien The most challenging issue in Biomanufacturing is the process constrained on minimal and maximal times for specific process. When molecules limitation versus time limitation assessing the relative incidence of the exhaustion between supply before death versus death before exhaustion of supply and highly depends on the quality of molecules. On the other hand, the stochasticity in market demand precludes a precise evolutionary matching of time allocation and productivity opportunity. This problem is modeled as a multiobjective Pareto optimality problem to minimize the cost per material and cost per time subject to constraints of accelerated evolutionary distribution of molecules. n WA67 West Bldg 105B Joint Session ISS/Practice Curated: Empirical Studies on Platforms Sponsored: Information Systems Sponsored Session Chair: Fujie Jin, Kelley School of Business, Indiana University, Bloomington, IN, 47401, United Statesö 1 - A Deep Learning Approach to Better Understanding of Hospital Quality Weiguang Wang, University of Maryland, 3330 Van Munching Hall, R.H. Smith Business School, College Park, MD, 20740, United States, Guodong (Gordon) Gao Hospitals are plagued with quality issues. It is well recognized that factors of providers and patients both contribute to quality issues, but how to separate their effects remains a challenge. We design a deep-learning based method that shows great promise in solving this problem. Firstly, word embedding techniques are adopted to vectorize all elements captured by EHRs. Then a one-to-one model is built using machine learning to link vector representations to hospital quality. Finally, a modification process to remove the effect of certain factors is performed to examine the impact of them on hospital quality. The proposed model is applied to real-world examples using the Florida provider data. 2 - Seeking a Reward or Helping the Entrepreneur-backers Response to a Low-price Probabilistic Choice in Reward-based Crowdfunding Alvin Zuyin Zheng, Temple University, 1801 N. Broad St, Pennsylvania, Philadelphia, PA, 19122, United States, Jing Gong, Paul Pavlou Crowdfunding aims to collect small investments from a large number of backers. Since crowdfunding projects typically attract a small number of backers, this study examines the role of the lottery in crowdfunding outcomes. Using a four-year dataset from a reward-based crowdfunding platform in China, we show that although the lottery does indeed help attract a higher number of backers for a project, it reduces the total money raised and the probability of reaching the funding goal. Mechanism analyses show the lottery incentivizes people who would otherwise not fund the project to become opportunistic backers, the lottery also cannibalizes prospective rewardees and donors. 3 - How Does Online Lending Influence Bankruptcy Filings? Evidence from a Natural Experiment Hongchang Wang, Georgia Institute of Technology, Atlanta, GA, United States, Eric Overby By providing relatively quick and easy access to credit, online lending platforms may help people overcome financial setbacks and/or refinance high-interest debt, thereby decreasing bankruptcy filings. On the other hand, these platforms may cause people to overextend themselves financially, leading to a “debt trap and increasing bankruptcy filings. To investigate the impact of online lending on bankruptcy filings, we leverage a natural experiment: Lending Club’s entry into different states at different times. Using coarsened exact matching and a difference-in-differences approach, we find that Lending Club’s entry increases bankruptcy filings by approximately 8%. 4 - Beauty and Counter-signaling in Two-sided Matching Markets: Evidence from a Randomized Field Experiment Lanfei Shi, University of Maryland, 3330 Van Munching Hall, College Park, MD, 20742, United States While online platforms place trust at the center, markets with few alternative trust-inducing signals face an even bigger challenge. In such cases, we find that phone verification, when making it optional and visible to others, plays a more significant and strategic role that works as an effective signaling device. We also identify interesting differential ex-ante opt-in decisions and ex-post impact of verification across two sides of the platform. We further discuss the underlying mechanism by applying state-of-the-art deep learning techniques to mine attractiveness using images. Moreover, upon verification users become more proactive, which also contributes to the increase in matching.

Big Data Modeling and Monitoring Sponsored: Quality, Statistics and Reliability Sponsored Session

Chair: Xiaochen Xian, UW-Madison, Madison, WI, 53705, United States Co-Chair: Kaibo Liu, UW-Madison, Madison, WI, 53706, United States 1 - Unsupervised Point Anomaly Detection Using Neighborhood Structure Assisted Non-negative Matrix Factorization Imtiaz Ahmed, Texas A. & M. University, College Station, TX, 77840, United States, Yu Ding, Xia Ben Hu Anomaly detection is unsupervised in nature and we can only rely on the neighborhood structure of a point for its evaluation. In this paper, we develop an anomaly detection procedure under the non-negative matrix factorization (NMF) framework. We incorporate the structural similarity information in the original NMF setting and propose a new approach called Neighborhood Structure Assisted NMF (NS-NMF). We argue that it will increase the anomaly detection capability of the regular NMF and provide numerical evidence in support of our claim. We also compare our method with other close developments in the same family and provide a detailed comparative evaluation on 20 benchmark data sets. 2 - Concept Drift Monitoring of Supervised Learning Models via Score Functions Kungang Zhang, Daniel Apley, Anh Bui Predictive models are trained on historical data and applied to new data. However, the true predictive relationship between the response and the predictors may change over time (aka, concept drift), rendering the fitted model obsolete. We propose to monitor such drift via multivariate control charting to detect changes in the mean of the score function, which is zero-mean if there is no drift. The score function is automatically produced in stochastic gradient optimization as the derivative of the log-likelihood. The method provides guidance on when the model should be updated for optimal prediction following a drift, as well as diagnostic information on the nature of the drift. 3 - A Density-based Index for Clustering Validation Behnam Tavakkol, Stockton University, Galloway, NJ, United States, Myong Kee Jeong, Susan Albin Clustering validity indices are the main tools of evaluating the quality of clusters. In this study, we develop a density-based clustering validity index. As opposed to most clustering validity indices that capture the characteristics of clusters by representative statistics, the proposed validity index performances well on clusters with arbitrary shapes. 4 - Spatiotemporal Modeling and Real-time Prediction of Origin-destination Traffic Demand Xiaochen Xian, UW-Madison, Madison, WI, 53705, United States, Kaibo Liu New technology has enabled new opportunities in a data-rich environment for the traffic research. To provide an accurate input for traffic planning, scheduling, and optimization problems, we propose a spatiotemporal modeling technique for the OD demand of the traffic network. In particular, the OD demand counts are handled by a Poisson log-normal model; and the mean and covariance matrix of the model are parametrized by the features of the traffic network. Then the method estimates the parameters using the EM algorithm and online predicts the traffic demand collectively considering both the spatial and temporal correlations. n WA69 West Bldg 106A Statistical Aspects of Computer Experiments and Stochastic Simulation Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Qiong Zhang, Richmond, VA, 23284, United States 1 - Asymmetric Kriging Emulator for Risk Measurement of Stochastic Simulation Qiong Zhang, 1015 Floyd Ave, Richmond, VA, 23284, United States Expectile is known to be the only risk measure that is both coherent and elicitable. Due to its good property, expectile recently attracts more and more attention in risk management. In stochastic simulation, it is often required to precisely estimate the system risk performance for a large number of input points. Therefore, it is particularly important to develop a statistical emulator for the expectiles of stochastic simulation. In this talk, I will introduce a new approach, called Asymmetric Kriging, to emulate the expectile risk measure of stochastic simulation. Our approach can be reduced to the state-of-art approach, stochastic Kriging, in emulating the mean performance measure.

435