Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SC68

n SC66 West Bldg 105A Business Value of Artificial Intelligence Sponsored: Artificial Intelligence Sponsored Session Chair: Zhongju Zhang, Arizona State University, Tempe, AZ, 85287, United States 1 - Social Learning in Prosumption: Evidence from a Randomized Field Experiment Tianshu Sun, University of Southern California, 3670 Trousdale Parkway, Bridge Hall, BRI 310B, Los Angeles, CA, 90089, United States, JaeHwuen Jung, Ravi Bapna, Joseph Golden A customer can now actively participate in the design for a wide range of products, as a ‘prosumer’. However, a major challenge in prosumption is the effort involved and ideas required. We combine a randomized experiment with machine learning to examine whether social learning, the act of showing the focal user creations made by other users, can improve users’ project creation and purchase decision. We find that under certain conditions, showing other users’ design can be highly effective. We further explore heterogeneity in treatment effect across image and user characteristics and find that three types of social learning mechanisms - idea learning, product learning, and self-efficacy - are at play. 2 - High-order Proximity Preserving Information Network Hashing Yong Ge, University of Arizona, 5776 S.Tiger Lily PL, Tucson, AZ, 85747, United States Information network embedding is an effective way for efficient graph analytics. We propose a MF-based Information Network Hashing (INH-MF) algorithm, to learn binary codes which can preserve high-order proximity. We also suggest Hamming subspace learning,which only updates partial binary codes each time, to scale up INH-MF. We finally evaluate INH-MF on four real-world information network datasets with respect to the tasks of node classification and node recommendation. The results demonstrate that INH-MF can perform significantly better than competing learning to hashmethods in both tasks, and surprisingly outperforms network embedding methods 3 - Crowdboosting: A Boosting-based Model for Crowd Opinions Aggregation Qianzhou Du, Virginia Tech, Blacksburg, VA, 24060, United States, Alan Gang Wang, Weiguo Fan Most existing crowd opinion aggregation studies employ rules to determine a weighting scheme for judges based on their past prediction performance. However, those rule-based methods often fail to determine the optimal weights and achieve the optimal prediction performance. We propose a new crowd opinion aggregation model, namely CrowdBoosting, which has a combining procedure for converting weak learners to strong ones and a mechanism of estimating the probabilities of event outcomes based on statistical learning. We empirically evaluate CrowdBoosting in comparison to four baseline methods. The results show that CrowdBoosting significantly outperforms all the baseline methods. 4 - Towards Better Learning from Crowd Labeling Junming Yin, University of Arizona, Management Information Systems Department, McClelland Hall, Room 430BB, Tucson, AZ, 85721, United States Microtask crowdsourcing has emerged as a cost-effective approach to obtaining large-scale labeled data in a wide range of applications. Crowdsourcing platforms such as MTurk and CrowdFlower provide an online marketplace where task requesters can submit a batch of microtasks, which a crowd of workers can then complete for a small monetary compensation. However, as the information provided by crowd workers can be prone to errors, significant effort is required to infer the true labels from such noisy labels supplied by a set of heterogeneous workers. Moreover, it would be very beneficial to identify (and then possibly filter out) those crowd workers with low reliabilities so as to foster the creation of a healthy and sustainable crowdsourcing ecosystem. Existing literature on learning from crowds has mainly focused on the single-label (i.e., binary and multi-class) setting, which prevents the application of microtask crowdsourcing to a wide range of business applications in which each item can be associated with multiple labels simultaneously. In this work, we consider the problem of learning from crowd labeling in the general multi-label setting, which includes previously studied single-label crowdsourcing problems as special cases. We propose a new Bayesian hierarchical model for the underlying annotation process of crowd workers, and introduce a mixture of Bernoulli distribution to capture the unknown label dependency. An efficient variational inference procedure is then developed to jointly infer ground truth labels, worker reliability in terms of sensitivity and specificity, and label dependency. Results based on extensive simulation experiments and a real-world MTurk experiment indeed confirm that the proposed approach outperforms other competing methods, highlighting the necessity to model both worker quality and label dependency when learning from crowdsourced multi-label annotations.

n SC67 West Bldg 105B Economics of IS Sponsored: Information Systems Sponsored Session Chair: Jingjing Li, University of Virginia

1 - An Empirical Evidence of Price Dispersion in Electronic Markets Jin Sik Kim, University of California-Irvine, 5102 Palo Verde Road, Irvine, CA, 92617, United States, Vijay C. Gurbaxani Theory predicts that the price of comparable product at online retailers would exhibit low price dispersion; yet, there is empirical evidence to the contrary. This paper investigates online price dispersion in two product types. We hypothesize that product types may influence price dispersion differently because of uncertainty to customers. Seller information may play a vital role in forming dispersed price. Lastly, we assume that search costs may lead to price dispersion. Our logistic regression results show the product with high uncertainty lead to higher price dispersion, some of the seller information increases price dispersion, and search costs are positively related to price dispersion. 2 - Are More Diverse Crowds Smarter? Jiayu Yao, Georgia Institute of Technology, Atlanta, GA, United States, Qiang Gao, Mingfeng Lin Does the diversity increase the wisdom of crowds? We examine the value of diversity in crowds within a market setting, specifically, the online financial market, utilizing a natural experiment on Prosper.com. 3 - Personalized Recommendations and Platform Monetization: Theory and Evidence from an App Portal Shengjun Mao, University of California-Irvine, Paul Merage School of Business SB1 3300, University of California, Irvine, Irvine, CA, 92697, United States, Sanjeev Dewan, Yi-Jen (Ian) Ho We examine app monetization on a mobile portal. Specifically, we study the click and conversion performance as a function of screen rank, app popularity, and app quality. We also apply natural language processing to investigate the app titles and descriptions in terms of the ambiguity, informativeness, and sentiment. We develop a bivariate probit model to jointly analyze click and conversion choices using a hierarchical Bayesian approach. Our results not only confirm the effects of ranking, popularity and quality but show the subjective text information matters. More importantly, our analyses help optimize the app placements to maximize the profits of the portal through personalization. n SC68 West Bldg 105C Technometrics Invited Session Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Alaa Elwany, Texas A&M University, College Station, TX, 77843, United States Co-Chair: Xiao Liu Co-Chair: Daniel Apley, Northwestern University, Evanston, IL, 60208- 3119, United States 1 - Bayesian Model Building from Small Samples of Disparate Data for Capturing In-plane Deviation in Additive Manufacturing Arman Sabbaghi, Purdue University, 150 N. University Street, West Lafayette, IN, 47907, United States, Qiang Huang, Tirthankar Dasgupta Quality control of shape deviation in additive manufacturing relies on statistical deviation models. However, resource constraints limit the manufacture of test shapes, and consequently impede the specification of models for new shapes. We present an adaptive Bayesian methodology that effectively combines in-plane deviation data and models for a small sample of previously manufactured, disparate shapes to aid in the model specification of in-plane deviation for new shapes. The power and simplicity of this method is demonstrated with case studies on in-plane deviation modeling for polygons and straight edges in free- form shapes using only data and models for cylinders and a single pentagon.

85