Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB68

3 - Detection of Prejudice from Social Media Streams Haimonti Dutta, University at Buffalo, 325P Jacobs Management Center, Buffalo, NY, 14260, United States Multicultural societies are characterized by frequent contact and communication between different social groups. Prejudices, or beliefs about a particular social group, affect the nature and quality of interactions with members of that group. It finds expression in social media platforms when a group of people express anger, resentment and dissent towards another. It is heightened during crises or threat. This paper develops a framework for detection of prejudice from social media streams using deep learning methods. Empirical results presented on a Twitter data stream show that our framework is capable of out-performing state-of-the- art batch systems for prejudice detection. 4 - Forecasting Collective Action using Social Media Rostyslav Korolov, Rensselaer Polytechnic Institute, 110 Eighth Street, Troy, NY, 12180, United States, David Mendonca, William A. Wallace We demonstrate the potential of online social media for large-scale behavioral research by developing a multi-disciplinary approach to predicting collective action using social media data. We report three case studies: one concerning charitable donations and two concerning social unrest. Through these case studies we demonstrate the utility of social media data for predictions of different kinds of collective action. We utilize Twitter data (approximately 20 million messages) to predict the volume of disaster response donations, and occurrences of protests. 5 - Predicting Gasoline Shortage in Florida During Irma using Tweets Abhinav Khare, University at Buffalo, Buffalo, New York, NY, United States, Rajan Batta, Qing He Shortage of supplies during a disaster is a common issue. Social media usage during disasters also surges .We developed a model to infer demand of supplies during disaster using Twitter and tested it to infer gasoline demand that occurred in Florida during Irma. We built a support vector machine model combined with n-grams and topic models to classify tweets regarding demand and modelled the arrival of tweets as a Space-TIme Poisson Process to infer the spatiotemporal demand. The model was tested on a database of 1.4 million tweets around Hurricane Irma in Florida. Our classification accuracy is high with f-score of 0.879 and the predicted demand correlates highly with the ground truth. n TB67 West Bldg 105B Platforms and Consumer Behavior Sponsored: Information Systems Sponsored Session Chair: Yuheng Hu, University of Illinois-Chicago, Chicago, IL, 60607, United States 1 - Can (s)he Code? Gender Bias in an Open Source Software Community Tingting Nian, University of California, Irvine, UCI, Irvine, CA, United States It is well-documented that women are underrepresented across all STEM fields and particularly in computing. Previous research identifies the negative stereotype about women’s abilities is one of the leading reason. In order to scientifically examine the issue of gender bias in online open source software communities, our study investigates whether and to what extent there exists a gender bias in the programming question-and-answer site, Stack Overflow. Our results reveal a net discrimination of 0.4 votes per month against female participants, statistically significant at the one percent level. We also found evidence suggestive of the statistical discrimination at work. 2 - Impact of Free Shipping Threshold on Different Channels: Evidence from an Online Retailer Fujie Jin, Kelley School of Business, Indiana University, 1309 E. Tenth Street, HH 4100, Bloomington, IN, 47401, United States, Fei Gao, Jianbin Li In this working project, we use a unique data from an online retailer to look at customer padding behavior, with respect to a shipping fee change. Our preliminary evidence suggest that customers using the PC channel are more likely to pad their shopping cart to meet the minimum shipping fee, compared with customers using the mobile channel. We suggest that difference in search cost across the two channels is a main contribute to such different user behavior. 3 - Visualizing Cybersecurity Risk Factors Arion Cheong, Rutgers University, Newark, NJ, United States, Won Gyun No, Soohyun Cho The destructive nature of cyber threats makes it crucial for stakeholders of a firm to understand the cybersecurity risks that the firm faces. To outline a firm’s known cybersecurity risks and identify its specific risks as compared to its industry peers, we utilize text mining to extract and analyze risk factor disclosures that firms report on Form 10-K. Since a comparison of cybersecurity risks through textual analysis is statistically infeasible, our approach visualizes each firm’s

cybersecurity risks as an image in the comparison process. This visualization enables stakeholders’ decision-making processes to easily access and compare cybersecurity risks. 4 - Fun Shopping - A Randomized Filed Experiment of Gamification Yi-Jen (Ian) Ho, Pennsylvania State University, 465 Business Building, Smeal College of Business, University Park, PA, 16802, United States, Siyuan Liu, Lei Wang In this research, we conduct a large-scale randomized field experiment at one of the largest Asian shopping mall to investigate the impact of gamification on customer engagement. The results shows gamification not only user engagement in terms of time spent and distance walked, but increase stores’ sales. We further benchmark this impact with the effect of traditional couponing. This study provides important implications on how firms can take advantages of gamification. 5 - The Effect of Search Costs and Stockouts on Consumer Search Behavior and Price Competition Xingyue (Luna) Zhang, University of Washington Tacoma, 621 Taylor Street, Tacoma, WA, 98402, United States, James Dearden, Yuliang Yao Consumers face various costs during shopping: the travel costs to visit stores and the search cost to examine product attributes once at a store. We build game- theoretic models and examine the interaction between store pricing and consumer search behavior with different costs. Our findings suggest that: 1) stores set equilibrium prices higher than marginal costs with the existence of the travel and search costs as well as stockout probabilities, 2) following product stockouts, stores are better off by giving up consumers when the attribute search cost is high, and 3) consumers are less likely to visit a store when the costs and/or stockout probabilities are high such that the prices are lower. n TB68 West Bldg 105C High Dimensional Data Analytics for Smart and Connected Systems Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Raed Al Kontar, University of Wisconsin-Madison, Madison, WI, 53706, United States Co-Chair: Chenang Liu, Blacksburg, VA, 24060, United States 1 - Minimizing Negative Transfer of Knowledge in Multivariate Gaussian Processes: A Scalable and Regularized Approach Raed Al Kontar, University of Wisconsin-Madison, Madison, WI, 53706, United States We propose a regularized and scalable modeling approach for the multivariate Gaussian process established using a convolution process. The key feature of our approach is the ability to minimize the negative transfer of knowledge between uncorrelated output through penalizing latent functions that facilitate information sharing. Statistical guarantees for the proposed method are studied and its advantageous features are demonstrated through numerical studies. 2 - A Multiplex Network Modeling Approach for Online Process Monitoring Chenang Liu, 845 Claytor Square, Blacksburg, VA, 24060, United States, Zhenyu Kong The objective of this research is to implement a new online monitoring method for complex systems using high dimensional sensing data. To achieve this objective, a multiplex network-based modeling approach is proposed in this study. The novelty of this method is to describe the sensing data using an effective multiplex network structure. Case studies in manufacturing and healthcare applications demonstrate that the proposed method can significantly enhance the monitoring sensitivity compared to the conventional methods. 3 - Deterministic and Stochastic Data Decomposition for Analytics of Complex Systems Xiaowei Yue, Georgia Institute of Technology, 755 Ferst Drive NW, ISYE, Atlanta, GA, 30332, United States, Jianjun Shi Data decomposition is an important step for high-dimensional data analytics of complex systems. This paper summarizes the key techniques for data decomposition, and separates them into two categories. One is deterministic decomposition, and the other is stochastic decomposition. The deterministic decomposition captures geometric or algebraic shape from the high-dimensional datasets directly, which is efficient for feature extraction and dimensionality reduction; while the stochastic decomposition provides probabilistic descriptions, and statistical distributions are estimated from the datasets.

299