Probabilistic Models for Scalable Knowledge Graph Construction

In the past decade, systems that extract information from millions of Internet documents have become commonplace. Knowledge graphs -- structured knowledge bases that describe entities, their attributes and the relationships between them -- are a powerful tool for understanding and organizing this vast amount of information. However, a significant obstacle to knowledge graph construction is the unreliability of the extracted information, due to noise and ambiguity in the underlying data or errors made by the extraction system and the complexity of reasoning about the dependencies between these noisy extractions. My dissertation addresses these challenges by exploiting the interdependencies between facts to improve the quality of the knowledge graph in a scalable framework. I introduce a new approach called knowledge graph identification (KGI), which resolves the entities, attributes and relationships in the knowledge graph by incorporating uncertain extractions from multiple sources, entity co-references, and ontological constraints. I define a probability distribution over possible knowledge graphs and infer the most probable knowledge graph using a combination of probabilistic and logical reasoning. Such probabilistic models are frequently dismissed due to scalability concerns, but my implementation of KGI maintains tractable performance on large problems through the use of hinge-loss Markov random fields, which have a convex inference objective. This allows the inference of large knowledge graphs using 4M facts and 20M ground constraints in 2 hours. To further scale the solution, I develop a distributed approach to the KGI problem which runs in parallel across multiple machines, reducing inference time by 90%. Finally, I extend my model to the streaming setting, where a knowledge graph is continuously updated by incorporating newly extracted facts. I devise a general approach for approximately updating inference in convex probabilistic models, and quantify the approximation error by defining and bounding inference regret for online models. Together, my work retains the attractive features of probabilistic models while providing the scalability necessary for large-scale knowledge graph construction. These models have been applied on a number of real-world knowledge graph projects, including the NELL project at Carnegie Mellon and the Google Knowledge Graph.

ASSESSING QUALITY IN HIGH-UNCERTAINTY MARKETS: ONLINE REVIEWS OF CREDENCE SERVICES

In economics of information theory, credence products are those whose quality is difficult or impossible for consumers to assess, even after they have consumed the product (Darby & Karni, 1973). This dissertation is focused on the content, consumer perception, and power of online reviews for credence services. Economics of information theory has long assumed, without empirical confirmation, that consumers will discount the credibility of claims about credence quality attributes. The same theories predict that because credence services are by definition obscure to the consumer, reviews of credence services are incapable of signaling quality. Our research aims to question these assumptions. In the first essay we examine how the content and structure of online reviews of credence services systematically differ from the content and structure of reviews of experience services and how consumers judge these differences. We have found that online reviews of credence services have either less important or less credible content than reviews of experience services and that consumers do discount the credibility of credence claims. However, while consumers rationally discount the credibility of simple credence claims in a review, more complex argument structure and the inclusion of evidence attenuate this effect. In the second essay we ask, “Can online reviews predict the worst doctors?” We examine the power of online reviews to detect low quality, as measured by state medical board sanctions. We find that online reviews are somewhat predictive of a doctor’s suitability to practice medicine; however, not all the data are useful. Numerical or star ratings provide the strongest quality signal; user-submitted text provides some signal but is subsumed almost completely by ratings. Of the ratings variables in our dataset, we find that punctuality, rather than knowledge, is the strongest predictor of medical board sanctions. These results challenge the definition of credence products, which is a long-standing construct in economics of information theory. Our results also have implications for online review users, review platforms, and for the use of predictive modeling in the context of information systems research.

Active Data Collection Techniques to Understand Online Scammers and Cybercriminals

Nigerian scam, also known as advance fee fraud or 419 scam, is a prevalent form of online fraudulent activity that causes financial loss to individuals and businesses. Nigerian scam has evolved from simple non-targeted email messages to more sophisticated scams targeted at users of classifieds, dating and other websites. Even though such scams are observed and reported by users frequently, the community’s understanding of Nigerian scams is limited since the scammers operate “underground”. To better understand the underground Nigerian scam ecosystem and seek effective methods to deter Nigerian scam and cybercrime in general, we conduct a series of active and passive measurement studies. Relying upon the analysis and insight gained from the measurement studies, we make four contributions: (1) we analyze the taxonomy of Nigerian scam and derive long-term trends in scams; (2) we provide an insight on Nigerian scam and cybercrime ecosystems and their underground operation; (3) we propose a payment intervention as a potential deterrent to cybercrime operation in general and evaluate its effectiveness; and (4) we offer active and passive measurement tools and techniques that enable in-depth analysis of cybercrime ecosystems and deterrence on them. We first created and analyze a repository of more than two hundred thousand user-reported scam emails, stretching from 2006 to 2014, from four major scam reporting websites. We select ten most commonly observed scam categories and tag 2,000 scam emails randomly selected from our repository. Based upon the manually tagged dataset, we train a machine learning classifier and cluster all scam emails in the repository. From the clustering result, we find a strong and sustained upward trend for targeted scams and downward trend for non-targeted scams. We then focus on two types of targeted scams: sales scams and rental scams targeted users on Craigslist. We built an automated scam data collection system and gathered large-scale sales scam emails. Using the system we posted honeypot ads on Craigslist and conversed automatically with the scammers. Through the email conversation, the system obtained additional confirmation of likely scam activities and collected additional information such as IP addresses and shipping addresses. Our analysis revealed that around 10 groups were responsible for nearly half of the over 13,000 total scam attempts we received. These groups used IP addresses and shipping addresses in both Nigeria and the U.S. We also crawled rental ads on Craigslist, identified rental scam ads amongst the large number of benign ads and conversed with the potential scammers. Through in-depth analysis of the rental scams, we found seven major scam campaigns employing various operations and monetization methods. We also found that unlike sales scammers, most rental scammers were in the U.S. The large-scale scam data and in-depth analysis provide useful insights on how to design effective deterrence techniques against cybercrime in general. We study underground DDoS-for-hire services, also known as booters, and measure the effectiveness of undermining a payment system of DDoS Services. Our analysis shows that the payment intervention can have the desired effect of limiting cybercriminals’ ability and increasing the risk of accepting payments.

Online Supplement to ‘Myopic Allocation Policy with Asymptotically Optimal Sampling Rate’

This document is the Online Supplement to ‘Myopic Allocation Policy with Asymptotically Optimal Sampling Rate,’ to be published in the IEEE Transactions of Automatic Control in 2017.

Flooded with Information from Social Media: Effects of Disaster Information Source and Visuals on Viewers' Cognitive, Affective, and Behavioral Responses

While a variety of crisis types loom as real risks for organizations and communities, and the media landscape continues to evolve, research is needed to help explain and predict how people respond to various kinds of crisis and disaster information. For example, despite the rising prevalence of digital and mobile media centered on still and moving visuals, and stark increases in Americans’ use of visual-based platforms for seeking and sharing disaster information, relatively little is known about how the presence or absence of disaster visuals online might prompt or deter resilience-related feelings, thoughts, and/or behaviors. Yet, with such insights, governmental and other organizational entities as well as communities themselves may best help individuals and communities prepare for, cope with, and recover from adverse events. Thus, this work uses the theoretical lens of the social-mediated crisis communication model (SMCC) coupled with the limited capacity model of motivated mediated message processing (LC4MP) to explore effects of disaster information source and visuals on viewers’ resilience-related responses to an extreme flooding scenario. Results from two experiments are reported. First a preliminary 2 (disaster information source: organization/US National Weather Service vs. news media/USA Today) x 2 (disaster visuals: no visual podcast vs. moving visual video) factorial between-subjects online experiment with a convenience sample of university students probes effects of crisis source and visuals on a variety of cognitive, affective, and behavioral outcomes. A second between-subjects online experiment manipulating still and moving visual pace in online videos (no visual vs. still, slow-pace visual vs. still, medium-pace visual vs. still, fast-pace visual vs. moving, slow-pace visual vs. moving, medium-pace visual vs. moving, fast-pace visual) with a convenience sample recruited from Amazon’s Mechanical Turk (mTurk) similarly probes a variety of potentially resilience-related cognitive, affective, and behavioral outcomes. The role of biological sex as a quasi-experimental variable is also investigated in both studies. Various implications for community resilience and recommendations for risk and disaster communicators are explored. Implications for theory building and future research are also examined. Resulting modifications of the SMCC model (i.e., removing “message strategy” and adding the new category of “message content elements” under organizational considerations) are proposed.

RACISM IN DIGITAL ERA: DEVELOPMENT AND INITIAL VALIDATION OF THE PERCEIVED ONLINE RACISM SCALE (PORS V1.0)

Racism continues to thrive on the Internet. Yet, little is known about racism in online settings and the potential consequences. The purpose of this study was to develop the Perceived Online Racism Scale (PORS), the first measure to assess people’s perceived online racism experiences as they interact with others and consume information on the Internet. Items were developed through a multi-stage process based on literature review, focus-groups, and qualitative data collection. Based on a racially diverse large-scale sample (N = 1023), exploratory and confirmatory factor analyses provided support for a 30-item bifactor model with the following three factors: (a) 14-item PORS-IP (personal experiences of racism in online interactions), (b) 5-item PORS-V (observations of other racial/ethnic minorities being offended), and (c) 11-item PORS-I (consumption of online contents and information denigrating racial/ethnic minorities and highlighting racial injustice in society). Initial construct validity examinations suggest that PORS is significantly linked to psychological distress.

AN EMPIRICAL ASSESSMENT OF USER ONLINE SECURITY BEHAVIOR: EVIDENCE FROM A UNIVERSITY

The ever-increasing number and severity of cybersecurity breaches makes it vital to understand the factors that make organizations vulnerable. Since humans are considered the weakest link in the cybersecurity chain of an organization, this study evaluates users’ individual differences (demographic factors, risk-taking preferences, decision-making styles and personality traits) to understand online security behavior. This thesis studies four different yet tightly related online security behaviors that influence organizational cybersecurity: device securement, password generation, proactive awareness and updating. A survey (N=369) of students, faculty and staff in a large mid-Atlantic U.S. public university identifies individual characteristics that relate to online security behavior and characterizes the higher-risk individuals that pose threats to the university’s cybersecurity. Based on these findings and insights from interviews with phishing victims, the study concludes with recommendations to help similat organizations increase end-user cybersecurity compliance and mitigate the risks caused by humans in the organizational cybersecurity chain.