Informs Annual Meeting 2017

TB31

INFORMS Houston – 2017

TB28

hackers. Identifying similar hackers can aid in proactive CTI by providing actionable intelligence.

350D Algorithmic Game Theory and Data Science

TB30

Invited: Auctions Invited Session Chair: Vasilis Syrgkanis, Microsoft Research, Cambridge, MA, 02139, United States, vasy@microsoft.com 1 - Oracle-efficient Online Learning and Applications to Auction Design Nika Haghtalab, Carnegie Mellon University, 5000 Forbes Avenue, CS Department, Pittsburgh, PA, 15213, United States, nhaghtal@cs.cmu.edu, Miroslav Dudik, Haipeng Luo, Robert Schapire, Vasilis Syrgkanis, Jennifer Wortman Vaughan We consider the fundamental problem of no-regret learning from experts and its applications to auction design. We consider designing computationally efficient no-regret algorithms given access to an oracle that finds the historically best- performing expert. We apply these results to an online auction framework, where the seller’s goal is to maximize its revenue by adaptively choosing the auction that is run on an online series of bid profiles. Our algorithms are general-purpose and, by design, can use little overhead to tap into existing tools that optimize auctions over historical data, thereby achieving online no-regret results without the need for building new specialized tools from scratch. 2 - Fairness Incentives for Myopic Agents Jamie Morgenstern, University of Pennsylvania, Philadelphia, PA, United States, jamiemmt.cs@gmail.com We consider schemes to incentivize myopic agents (such as Airbnb landlords, who may emphasize short-term profits and property safety) to treat arriving clients fairly, in order to prevent overall discrimination against individuals or groups. Our notion of fairness asks that more qualified individuals are never (probabilistically) preferred over less qualified ones [Joseph et al]. We investigate whether it is possible to design inexpensive {subsidy} or payment schemes for a principal to motivate myopic agents to play fairly in all or almost all rounds. We show both positive and negative results in the classic and linear bandit settings by upper and lower bounding the cost of fair subsidy schemes. 3 - Deep Learning for Predicting Human Strategic Behavior James R. Wright, Microsoft Research, New York, NY, United States, jrwright@microsoft.com, Jason Hartford, Kevin Leyton-Brown Predicting the behavior of humans in strategic settings is an important problem in many domains. Most existing work either assumes that participants are perfectly rational, or attempts to directly model each participant’s cognitive processes based on insights from cognitive psychology and experimental economics. We present an alternative, a deep learning approach that automatically performs cognitive modeling without relying on such expert knowledge. We introduce a novel architecture that allows a single network to generalize across different input and output dimensions, and show that it significantly outperforms the previous state of the art, which relies on expert-constructed features. 350E Artificial Intelligence for Social Media Applications Sponsored: Artificial Intelligence Sponsored Session Chair: Sagar Samtani, sagars@email.arizona.edu 1 - Aspect Extraction and Polarity Classification of Reviews Based on Deep Neural Network Czang Yeob Kim, Korea University, Seoul, Korea, Republic of, czangyeob@korea.ac.kr, Pilsung Kang Consumers buy products and leave opinions online. Companies need to analyze the voice of consumers online to understand the response of the market. Consumer reviews can be analyzed with sentimental analysis at the aspect-level to gain insights for new products by reflecting the consumer’s response to current products’ functionality. We propose a deep neural network-based aspect extraction and polarity classification model. The aspect extraction model identifies the product feature the review is referring to, whereas polarity classification determines the review is whether positive or negative to the identified feature. The proposed model is verified based on Amazon’s review data. 2 - Semantics in Hacker Forums TB29

350F Social Media Analytics I Invited: Social Media Analytics Invited Session Chair: Shu He, University of Connecticut, Storrs, CT, 06268, United States, shuhe@utexas.edu 1 - Online to Offline: Impact of Social Advertising on Pro-social Behavior Cenying Yang, ycy_tracy@utexas.edu, Shun-Yang Lee, Andrew B. Whinston We attempt to link the social advertising to offline pro-social behavior. We launch advertising campaign on Facebook and test its effectiveness on actual pro-social behavior in real world. We run field experiments under the context of political voting, blood donation and HPV vaccine shot. Psychological mechanisms that could potentially drive people to behave pro-socially are explored. 2 - Unraveling the “Social” in Social Norms: the Conditioning Effect of User Connectivity Che-Wei Liu, University of Maryland, Robert H. Smith School of We studied the effect of social norms on users’ goal setting and goal attainment behavior. We find that while social influence increases the rate of goal setting, strikingly, we also observe an undesirable effect in that such influence leads to lower rates of goal attainment. Our examination of heterogeneous treatment effects shows that individuals with higher levels of social connectivity are the most susceptible. Further analysis suggests that social norms motivate these runners with high social connectivity to set up a goal beyond their own capabilities, resulting in low goal attainment. Our findings have important implications for the design of interventions in mobile health technologies. 3 - Impacts of Online Diaries on Sales of Credence Goods: Evidence from a Cosmetic Surgery Platform Hongfei Li, University of Connecticut, School of Business, Credence goods are defined as products whose quality is difficult to evaluate. While online reviews have found to be effective in driving sales of search and experience goods, little is known about its effectiveness on credence goods. Using data from a cosmetic surgery platform, we investigate the impact of a novel form of online reviews, i.e. diaries, on the sales of credence goods. We also examine how the effects are moderated by the risk of surgeries. We focus on two diary features: number of images and post duration. We find the number of images has a positive impact on sales and the effects are stronger for more risky surgeries. These findings shed some light on the online marketing of credence goods. 351A Financial Engineering Contributed Session Chair: Mingying Song, The University of Hong Kong, Hong Kong, mysong@hku.hk 1 - An Efficient Simulation Method for Normal Stochastic Volatility Model Jaehyuk Choi, Peking University HSBC Business School, Shenzhen, China, jaehyuk@phbs.pku.edu.cn, Byoung Ki Seo Normal, as opposed to lognormal, distribution is often suitable for modeling some financial observables which allow negative values. This study concerns a class of normal stochastic volatility model which includes the zero beta case of the popular Stochastic-Alpha-Beta-Rho (SABR) model. We provide an exact one-step simulation method and an efficient integration method for vanilla option pricing. A special case of the class admits a closed-form distribution, which can be used for modeling and simulating leptokurtic distributions. Business,, College Park, MD, 20742, United States, cwliu@rhsmith.umd.edu, Guodong Gao, Ritu Agarwal 2100 Hillside Road, Storrs, CT, 06269, United States, hongfei.li@uconn.edu, Gang Wang, Jing Peng, Xue Bai TB31

Sandeep Suntwal, University of Arizona, Tucson, AZ, United States, sandeepsuntwal@email.arizona.edu

Cyberattacks are a high impact technological risk. Cyber threat intelligence (CTI) is the process of understanding the threats to an organization based on available data points focusing on understanding key threat actors and relevant threat vectors. In this talk, we discuss a framework to model online hacker forums into Heterogeneous Information Networks (HIN) and identify semantically similar

313