Shawndra Hill

Research

Research Interests: data mining/knowledge discovery in databases, design science and information systems, dynamic networks, machine learning, network-based marketing, statistical relational learning

Current Project: The Social TV Lab

Year	Author	Title	Journal / Conference
2012	Adrian Benton, John H. Holmes, S. Hill, Annie Chung, Lyle Ungar	Medpie: an information extraction package for medical message board posts We have developed medpie, a software package for preparing medical message board corpora and extracting patient mentions and statistics for drugs, herbs, and adverse effects experienced from them. The package is divided into web-crawling, HTML-cleaning, de-identification, and information extraction modules. It also includes a sample controlled vocabulary of drugs, herbs, and adverse effect terms.	Bioinformatics 28(5): 743 (2012) (software release note)
2011	Shawndra Hill, Jun Mao, Lyle Ungar, Sean Hennessy, Charles E. Leonard, John H. Holmes	The Rise of H1N1 and Herbal Supplements As the incidence of H1N1 increases the lay public may turn to the internet for information about natural supplements for prevention and treatment. Objectives: To identify and characterize websites that provide information about herbal and natural supplements with information about H1N1, and to examine trends in the public's behavior in searching for information about supplement use in preventing or treating H1N1	Journal of Medical Internet Research, 14(2)
2011	Adrian Benton, Lyle Ungar, Shawndra Hill, Sean Hennessy, Jun Mao, Annie Chung, Charles E. Leonard, John H. Holmes	Identifying potential adverse effects using the web: A new approach to medical hypothesis generationMedical message boards are online resources where users with a particular condition exchange information, some of which they might not otherwise share with medical providers. Many of these boards contain many posts and contain patient opinions and experiences that would be potentially useful to clinicians and researchers. We present an approach that is able to collect a corpus of medical message board posts, de-identify the corpus, and extract information on potential adverse drug effects discussed by users. Using a corpus of posts to breast cancer message boards, we identified drug event pairs using co-occurrence statistics. We then compared the identified drug event pairs with adverse effects listed on the package labels of tamoxifen, anastrazole, exemestane, and letrozole. Of the pairs identified by our system, 75-80% were documented on the drug labels. Some pairs may represent previously unidentified adverse drug effects.	Journal of Biomedical Informatics, 44(6): 989-996
2011	Shawndra Hill, Noah Ready-Campbell	Expert Stock Picker: The Wisdom of (the Experts in) CrowdsThe phrase "the wisdom of crowds" suggests that good verdicts can be achieved by averaging the opinions and insights of large, diverse groups of people who possess varied types of information. Online user-generated content enables researchers to view the opinions of large numbers of users publicly. These opinions, in the form of reviews and votes, can be used to automatically generate remarkably accurate verdicts—collective estimations of future performance—about companies, products, and people on the Web to resolve very tough problems. The wealth and richness of user-generated content may enable firms and individuals to aggregate consumer-think for better business understanding. Our main contribution, here applied to user-generated stock pick votes from a widely used online financial newsletter, is a genetic algorithm approach that can be used to identify the appropriate vote weights for users based on their prior individual voting success.	International Journal of Electronic Commerce, 15(3): 73
2011	Getachew Berhan, Tsegaye Tadesse, Solomon Atnafu, Shawndra Hill	Using Satellite Images or Drought Monitoring: A Knowledge Discovery ApproachThe main objective of this research was to develop a new concept and approach to extracting knowledgefrom satellite imageries for near real-time drought monitoring in areas of Ethiopia that are experiencingfood insecurity. The near real-time data downloaded from the Atlantic Bird satellite were used to producethe drought spatial distribution. Our results showed that approximately 40% of the observed areasexhibited negative deviation. In this study, the possibility of using the near real-time spatio-temporal Meteosat Second Generation (MSG) data for drought monitoring in food insecure areas of Ethiopia wastested, and promising results were obtained. The output of this research is expected to assist decisionmakers in taking timely and appropriate action in order to save millions of lives in drought-affected areas.	Journal of Strategic Innovation and Sustainability, Vol. 7, No. 1, Spring 2011
2011	Adrian Benton, Shawndra Hill, Lyle Ungar, Annie Chung, John Holmes	A System for De-Identifying Medical Message Board TextThere are millions of public posts to medical message boards by users seeking support and information on a wide range of medical conditions. It has been shown that these posts can be used to gain a greater understanding of patients' experiences and concerns. As investigators continue to explore large corpora of medical discussion board data for research purposes, protecting the privacy of the members of these online communities becomes an important challenge that needs to be met. Extant entity recognition methods used for more structured text are not sufficient because message posts present additional challenges: the posts contain many typographical errors, larger variety of possible names, terms and abbreviations specific to Internet posts or a particular message board, and mentions of the authors' personal lives. The main contribution of this paper is a system to de-identify the authors of message board posts automatically, taking into account the aforementioned challenges. We demonstrate our system on two different message board corpora, one on breast cancer and another on arthritis. We show that our approach significantly outperforms other publicly available named entity recognition and de-identification systems, which have been tuned for more structured text like operative reports, pathology reports, discharge summaries, or newswire.	BMC Bioinformatics, 12(3):73
2006	Shawndra Hill, Deepak Agarwal, Robert Bell, Chris Volinsky	Building an Effective Representation for Dynamic NetworksA dynamic network is a special type of network composed of connected transactors which have repeated evolving interaction. Data on large dynamic networks such as telecommunications networks and the Internet are pervasive. However, representing dynamic networks in a manner that is conducive to efficient large-scale analysis is a challenge. In this article, we represent dynamic graphs using a data structure introduced in an earlier article. We advocate their representation because it accounts for the evolution of relationships between transactors through time, mitigates noise at the local transactor level, and allows for the removal of stale relationships. Our work improves on their heuristic arguments by formalizing the representation with three tunable parameters. In doing this, we develop a generic framework for evaluating and tuning any dynamic graph. We show that the storage saving approximations involved in the representation do not affect predictive performance, and typically improve it. We motivate our approach using a fraud detection example from the telecommunications industry, and demonstrate that we can outperform published results on the fraud detection task. In addition, we present a preliminary analysis on Web logs and e-mail networks.	Journal of Computational & Graphical Statistics, Vol. 15, No. 3, pp. 584-608 [Bibtex]
2006	Shawndra Hill, Foster Provost, Chris Volinsky	Network-based Marketing: Identifying Likely Adopters via Consumer NetworksNetwork-based marketing refers to a collection of marketing techniques that take advantage of links between consumers to increase sales. We concentrate on the consumer networks formed using direct interactions (e.g., communications) between consumers. We survey the diverse literature on such marketing with an emphasis on the statistical methods used and the data to which these methods have been applied. We also provide a discussion of challenges and opportunities for this burgeoning research topic. Our survey highlights a gap in the literature. Because of inadequate data, prior studies have not been able to provide direct, statistical support for the hypothesis that network linkage can directly affect product/service adoption. Using a new data set that represents the adoption of a new telecommunications service, we show very strong support for the hypothesis.	Statistical Science, Vol. 21, No. 2, pp. 256-276 [Bibtex]
2005	Shawndra Hill, Abraham Bernstein, Foster Provost	Toward Intelligent Assistance for a Data Mining Process: An Ontology--Based Approach for Cost-Sensitive ClassificationA data mining (DM) process involves multiple stages. A simple, but typical, process might include preprocessing data, applying a data mining algorithm, and postprocessing the mining results. There are many possible choices for each stage, and only some combinations are valid. Because of the large space and nontrivial interactions, both novices and data mining specialists need assistance in composing and selecting DM processes. Extending notions developed for statistical expert systems we present a prototype Intelligent Discovery Assistant (IDA), which provides users with 1) systematic enumerations of valid DM processes, in order that important, potentially fruitful options are not overlooked, and 2) effective rankings of these valid processes by different criteria, to facilitate the choice of DM processes to execute. We use the prototype to show that an IDA can indeed provide useful enumerations and effective rankings in the context of simple classification processes. We discuss how an IDA could be an important tool for knowledge sharing among a team of data miners. Finally, we illustrate the claims with a demonstration of cost-sensitive classification using a more complicated process and data from the 1998 KDDCUP competition.	IEEE Transactions on Knowledge and Data Engineering, Vol. 17, Iss. 4, pp. 503-518 [Bibtex]
2003	Shawndra Hill, Foster Provost	The Myth of the Double-blind Review? Author Identification Using Only CitationsPrior studies have questioned the degree of anonymity of the double-blind review process for scholarly research articles. For example, one study based on a survey of reviewers concluded that authors often could be identified by reviewers using a combination of the author's reference list and the referee's personal background knowledge. For the KDD Cup 2003 competition's "Open Task," we examined how well various automatic matching techniques could identify authors within the competition's very large archive of research papers. This paper describes the issues surrounding author identification, how these issues motivated our study, and the results we obtained. The best method, based on discriminative self-citations, identified authors correctly 40-45% of the time. One main motivation for doubleblind review is to eliminate bias in favor of well-known authors. However, identification accuracy for authors with substantial publication history is even better (60% accuracy for the top-10% most prolific authors, 85% for authors with 100 or more prior papers).	SIGKDD Explorations, Vol. 5, Iss. 2, pp. 179-184 [Bibtex]
2012	Adrian Benton, Shawndra Hill	The Spoiler Effect?: Designing Social TV Content That Promotes Ongoing WOM Television shows are now instigating online social interactions between viewers by requesting viewers, as part of the first broadcasts, to participate in simultaneous discussions about shows. In this work, we analyze and quantify the effect of strategies used in an American reality singing show to engage viewers. We show three main results: 1) when specific tweet messages are posted on the TV screen during the show, the tweets are much more likely to be discussed by viewers than messages created by the same online social media users during the show that are not posted on the screen; 2) there is a difference between social media engagement response to specific tweet messages and more general social media content-- specific tweets with new information result in more engagement on the first show airing; and 3) there is a difference in overall response to specific social media tweet messages on the West Coast of the US since results by the time of airing have been spoiled.	CIST in Conjunction with INFORMS
2012	Shawndra Hill, Adrian Benton	Social TV: Analyzing the Impact of Social Media Strategies on Viewer Engagement “Social TV” is a term that broadly describes the online social interactions that occur between viewers while watching television. Television shows are now instigating online social interactions between viewers by requesting viewers, as part of the first broadcasts, to engage in simultaneous discussions about the shows. Examples include explicitly asking users to follow and comment on topics on the popular social media site, Twitter. A more passive approach used by TV shows displays messages from social media users on the TV screen during broadcasts. While many shows now include social media in their broadcasts, very little is known about the effectiveness of these strategies on engagement and viewership. In this work, we analyze features of the social media strategy of an American reality singing show. We show that when messages are posted on the TV screen during the show, they are much more likely to be discussed by viewers than messages created by the same online social media users during the show that are not posted on the screen, indicating that this strategy works. We also show that when a prompt is given to use a specific term in discussions online, viewers are more likely to remain engaged with the show on social media. We believe this is the first scholarly work to link real-time TV posts to realtime TV engagement response. Our contribution is a set of results, which show that prompting viewers during the show with social media posts on screen results in increased and prolonged engagement during the show.	The Second SOMA Workshop: Social Media Analytics
2012	Shawndra Hill, Aman Nalavade, Adrian Benton	Social TV: Real-Time Social Media Response to TV Advertising We link the content of real-time social media response to characteristics of the Super Bowl 2012 TV advertisements, including their advertisers' social media strategies and the categories of the products being advertised. We analyze millions of social media posts about approximately forty-five advertisements during the 2012 Super Bowl. While there are many studies that focus on the popularity of and sentiment of response to Super Bowl ads, we believe ours is the first to go deeper and analyze the textual "content" to understand better the factors contributing to specific types of social media responses among a number of TV advertising dimensions. In this study, we show that the level of online consumer engagement , measured by attracting new followers on Twitter during the Superbowl, can be linked to whether the brand that advertised had a social media strategy or not In addition, we show that sentiment response, a measure often use to quantify the effectiveness of TV ads, varies by demographic category.	ADKDD 2012: The Sixth International Workshop on Data Mining for Online Advertising and Internet Economy
2012	Shawndra Hill, Adrian Benton	Social TV: Linking Social Media Buzz to Sales	Conditionally Accepted ICIS (International Conference on Information Systems)
2011	Shawndra Hill, Benton A, Ungar L, Macskassy S, Chung A, Holmes JH	A Cluster-based method for isolating influence on TwitterThis paper demonstrates a cluster-based method to isolate influence in social networkbased observational data, where "influence" is defined to mean that one person posts about a topic online and a second person posts about the same topic because he or she read the first post. Isolating influence in observational data is difficult, because we may observe that connected people discuss the same topic in proximate periods for reasons other than influence, including homophily–connected people are similar–and exogenous shock; they may have learned of the topic from some external source. We employ a matched sample estimation technique that has been used in the past to measure influence by controlling for demographic and usage based homophily, and add to the matching scheme a cluster ID. Our contribution is two-fold: First, we provide preliminary evidence that social network-based clusters capture homophily, indicating that a network-based attribute approach may not only capture homophily but also may be used in lieu of using demographic attributes for matching similar users in scenarios when privacy preservation is a concern. Second, we show that by adding a network position attribute, a cluster ID, when matching similar users, we can isolate influence better. We believe that our approach to isolate influence can have a broad impact on problems where social networks and associated behaviors can be observed over time.	Workshop on Information Technologies and Systems
2011	Getachew Berhan, Shawndra Hill, Tsegaye Tadesse, Solomon Atnafu	Geographic Information Systems and Geo-statistics for Modeling and Mapping Endangered Species: A Case Study in Bonga Forest of Ethiopia	Global Information Technology Management Association
2010	Adrian Benton, Shawndra Hill, Lyle Ungar, Annie Chung, John Holmes	A System for De-Identifying Medical Message Board TextThere are millions of public posts to medical message boards by users seeking support and information on a wide range of medical conditions. It has been shown that these posts can be used to gain a greater understanding of patients' experiences and concerns. As investigators continue to explore large corpora of medical discussion board data for research purposes, protecting the privacy of the members of these online communities becomes an important challenge that needs to be met. Extant entity recognition methods used for more structured text are not sufficient because message posts present additional challenges: the posts contain many typographical errors, larger variety of possible names, terms and abbreviations specific to Internet posts or a particular message board, and mentions of the authors' personal lives. The main contribution of this paper is a system to de-identify the authors of discussion board posts automatically, taking into account the aforementioned challenges. We demonstrate our system on two different message board corpora, one on breast cancer and another on arthritis. We show that our approach significantly outperforms other publicly available de-identification systems, which have been tuned for more structured text like operative reports, pathology reports and discharge summaries. Our software will be available for download as open source code in the near future.	International Conference on Machine Learning Applications
2010	Getachew Berhan, Tsegaye Tadesse, Solomon Atnafu, Shawndra Hill	Knowledge Discovery from Satellite Images for Drought Monitoring: A Case Study in Ethiopia	Global Information Technology Management Association World Conference
2010	Justin Vastola, Kobi Abayomi, Shawndra Hill	Statistics for Re-Identification Signatures in Networked Data	Workshop on Information Networks
2010	Shawndra Hill, Getachew Berhan, Anita Banser, Nathan Eagle	Reality Mining AfricaCellular phones can be used as mobile sensors, continuously logging users' behavior including movement, communication and proximity to others. While it is well understood that data generated from mobile phones includes a record of phone calls, there are also more sophisticated data types, such as Bluetooth or cell tower proximity logging, which reveal movement patterns and day-to-day human interactions. We explore the possibility of using mobile phone data to compare movement and communication patterns across cultures. The goal of this proof-of-concept study is to quantify behavior in order to compare different populations. We compare our ability to predict future calling behavior and movement patterns from the cellular phone data of subjects in two distinct groups: a set of university students at MIT in the United States and the University of Nairobi in Kenya. In addition, we show how Bluetooth data may be used to estimate the diffusion of an airborne pathogen outbreak in the different populations.	AAAI Symposium on Artificial Intelligence for Development [Bibtex]
2010	Mariye Yigzaw, Shawndra Hill, Anita Banser, Lemma Lessa	Using Data Mining to Combat Infrastructure Inefficiencies: The Case of Predicting Non-payment for Ethiopian TelecomData mining and machine learning technologies for business applications have evolved over the past two decades, and are regularly applied in contemporary organizations to everything from manufacturing to online advertising in fields ranging from health care to motor racing. Unfortunately, data mining techniques are not applied as often to problems in the developing world. Despite the fact that some industries, such as banks, airlines, courts, and telecommunications firms, necessitate data storage as part of their business process. We argue that data mining could be used to reduce infrastructure inefficiencies, which is one of the largest problems faced by Africa. We demonstrate that we can potentially reduce the infrastructure inefficiency of the Ethiopian telecommunications industry by ranking customers according to their likelihood of nonpayment using a data mining approach.	AAAI Symposium on Artificial Intelligence for Development
2010	Tibebe Besheh, Shawndra Hill	Mining Road Traffic Accident Data to Improve Road Safety: The Role of Road-related Factors on Accident Severity in EthiopiaRoad traffic accidents (RTAs) are a major public health concern, resulting in an estimated 1.2 million deaths and 50 million injuries worldwide each year. In the developing world, RTAs are among the leading cause of death and injury; Ethiopia in particular experiences the highest rate of such accidents. Thus, methods to reduce accident severity are of great interest to traffic agencies and the public at large. In this work, we applied data mining technologies to link recorded road characteristics to accident severity in Ethiopia, and developed a set of rules that could be used by the Ethiopian Traffic Agency to improve safety.	AAAI Symposium on Artificial Intelligence for Development
2009	Shawndra Hill and Akash Nagle	Social Network Signatures: A Framework for Re‐Identification in Networked DataData on large dynamic social networks, such as telecommunications networks and the Internet, are pervasive. However, few methods conducive to efficient large-scale analysis exist. In this paper, we focus on the task of re-identification. Re-identification in the context of dynamic networks is a matching problem that involves comparing the behavior of networked entities across two time periods. Prior research has reported success in the domains of e-mail alias detection, author attribution, and identifying fraudulent consumers in the telecommunications industry. In this work, we address the question of "why are we able to re-identify entities on real world dynamic networks?" Our contribution is two-fold. First, we address the challenge of scale with a framework for matching that does not require pairwise comparisons to ascertain the similarity scores between networked entities. Second, we show our method is robust against missing links but less tolerant to noise. Using our framework, we provide a performance estimate for re-identification on networks based solely on their degree distribution and dynamics. This work has significant implications for re-identification problems where scale is a challenge as well as for problems where false negatives (e.g., when fraudulent consumers are not labeled as fraudulent) cannot be observed.	International Conference on Computational Aspects of Social Networks [Bibtex]
2007	Shawndra Hill, Foster Provost, Chris Volinsky	Learning and Inference in Massive Social NetworksResearchers and practitioners increasingly are gaining access to data on explicit social networks. For example, telecommunications and technology firms record data on consumer networks (via phone calls, emails, voice-over-IP, instant messaging), and social-network portal sites such as MySpace, Friendster and Facebook record consumer-generated data on social networks. Inference for fraud detection [5, 3, 8], marketing [9], and other tasks can be improved with learned models that take social networks into account and with collective inference [12], which allows inferences about nodes in the network to affect each other. However, these socialnetwork graphs can be huge, comprising millions to billions of nodes and one or two orders of magnitude more links.	5th International Workshop on Mining and Learning with Graphs [Bibtex]
2005	Shawndra Hill, Deepak Agarwal, Robert Bell, and Chris Volinsky	Tuning Representations of Dynamic Network DataA dynamic network is a special type of network which is comprised of connected transactors which have repeated evolving interaction. Data on large dynamic networks such as telecommunications networks and the Internet are pervasive. However, representing dynamic networks in a manner that is conducive to effcient large-scale analysis is a challenge. In this paper, we represent dynamic graphs using a data structure introduced by Cortes et. al. [3]. Our work improves on their heuristic arguments by formalizing the representation with three tunable parameters. In doing this, we develop a generic framework for evaluating and tuning any dynamic graph. We show that the storage saving approximations involved in the representation do not affect predictive performance, and typically improve it. We motivate our approach using a fraud detection example from the telecommunications industry, and demonstrate that we can outperform published results on the fraud detection task.	3rd International Workshop on Link Discovery
2003	Shawndra Hill	Social Network Relational Vectors for Anonymous Identity MatchingAnonymous fraudulent behavior can generate substantial financial burden and inconvenience. Moreover, the recent threat of terrorist infiltra-tion to both business and government has yielded heightened interest in anonymous iden-tity matching (AIM). Most applications of AIM require sophisticated methods robust to issues such as deliberate variation in identity attributes, missing data, and multi-source data corpora. We consider relational social network behavior, eliminating the reliance on personal identifiable data for identity matching. In particular, we consider problems that can be characterized by personal communication networks. We evaluate a proposed implementation of a social network vector-space relational model for AIM on Cite-Seer, a research publication citation database.	International Joint Conference on Artificial Intelligence
2002	Abraham Bernstein, Scott Clearwater, Shawndra Hill, Claudia Perlich, Foster Provost	Discovering Knowledge from Relational Data Extracted from Business NewsThousands of business news stories (including press releases, earnings reports, general business news, etc.) are released each day. Recently, information technology advances have partially automated the processing of documents, reducing the amount of text that must be read. Current techniques (e.g., text classification and information extraction) for full-text analysis for the most part are limited to discovering information that can be found in single documents. Often, however, important information does not reside in a single document, but in the relationships between information distributed over multiple documents. This paper reports on an investigation into whether knowledge can be discovered automatically from relational data extracted from large corpora of business news stories. We use a combination of information extraction, network analysis, and statistical techniques. We show that relationally interlinked patterns distributed over multiple documents can indeed be extracted, and (specifically) that knowledge about companies' interrelationships can be discovered. We evaluate the extracted relationships in several ways: we give a broad visualization of related companies, showing intuitive industry clusters; we use network analysis to ask who are the central players, and finally, we show that the extracted interrelationships can be used for important tasks, such as for classifying companies by industry membership.	8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Bibtex]

Year

Author

Title

Journal / Conference

2012

Adrian Benton, John H. Holmes, S. Hill, Annie Chung, Lyle Ungar

Medpie: an information extraction package for medical message board posts We have developed medpie, a software package for preparing medical message board corpora and extracting patient mentions and statistics for drugs, herbs, and adverse effects experienced from them. The package is divided into web-crawling, HTML-cleaning, de-identification, and information extraction modules. It also includes a sample controlled vocabulary of drugs, herbs, and adverse effect terms.

Bioinformatics 28(5): 743 (2012) (software
release note)

2011

Shawndra Hill, Jun Mao, Lyle Ungar, Sean Hennessy, Charles E. Leonard, John H. Holmes

The Rise of H1N1 and Herbal Supplements As the incidence of H1N1 increases the lay public may turn to the internet for information about natural supplements for prevention and treatment.
Objectives: To identify and characterize websites that provide information about herbal and natural supplements with information about H1N1, and to examine trends in the public's behavior in searching for information about supplement use in preventing or treating H1N1

Journal of Medical Internet Research, 14(2)

2011

Adrian Benton, Lyle Ungar, Shawndra Hill, Sean Hennessy, Jun Mao, Annie Chung, Charles E. Leonard, John H. Holmes

Identifying potential adverse effects using the web: A new approach to medical hypothesis generationMedical message boards are online resources where users with a particular condition exchange information, some of which they might not otherwise share with medical providers. Many of these boards contain many posts and contain patient opinions and experiences that would be potentially useful to clinicians and researchers. We present an approach that is able to collect a corpus of medical message board posts, de-identify the corpus, and extract information on potential adverse drug effects discussed by users. Using a corpus of posts to breast cancer message boards, we identified drug event pairs using co-occurrence statistics. We then compared the identified drug event pairs with adverse effects listed on the package labels of tamoxifen, anastrazole, exemestane, and letrozole. Of the pairs identified by our system, 75-80% were documented on the drug labels. Some pairs may represent previously unidentified adverse drug effects.

Journal of Biomedical Informatics, 44(6): 989-996

2011

Shawndra Hill, Noah Ready-Campbell

Expert Stock Picker: The Wisdom of (the Experts in) CrowdsThe phrase "the wisdom of crowds" suggests that good verdicts can be achieved by averaging the opinions and insights of large, diverse groups of people who possess varied types of information. Online user-generated content enables researchers to view the opinions of large numbers of users publicly. These opinions, in the form of reviews and votes, can be used to automatically generate remarkably accurate verdicts—collective estimations of future performance—about companies, products, and people on the Web to resolve very tough problems. The wealth and richness of user-generated content may enable firms and individuals to aggregate consumer-think for better business understanding. Our main contribution, here applied to user-generated stock pick votes from a widely used online financial newsletter, is a genetic algorithm approach that can be used to identify the appropriate vote weights for users based on their prior individual voting success.

International Journal of Electronic Commerce, 15(3): 73

2011

Getachew Berhan, Tsegaye Tadesse, Solomon Atnafu, Shawndra Hill

Using Satellite Images or Drought Monitoring: A Knowledge Discovery ApproachThe main objective of this research was to develop a new concept and approach to extracting knowledgefrom satellite imageries for near real-time drought monitoring in areas of Ethiopia that are experiencingfood insecurity. The near real-time data downloaded from the Atlantic Bird satellite were used to producethe drought spatial distribution. Our results showed that approximately 40% of the observed areasexhibited negative deviation. In this study, the possibility of using the near real-time spatio-temporal Meteosat Second Generation (MSG) data for drought monitoring in food insecure areas of Ethiopia wastested, and promising results were obtained. The output of this research is expected to assist decisionmakers in taking timely and appropriate action in order to save millions of lives in drought-affected areas.

Journal of Strategic Innovation and Sustainability, Vol. 7, No. 1, Spring 2011

2011

Adrian Benton, Shawndra Hill, Lyle Ungar, Annie Chung, John Holmes

A System for De-Identifying Medical Message Board TextThere are millions of public posts to medical message boards by users seeking support and information on a wide range of medical conditions. It has been shown that these posts can be used to gain a greater understanding of patients' experiences and concerns. As investigators continue to explore large corpora of medical discussion board data for research purposes, protecting the privacy of the members of these online communities becomes an important challenge that needs to be met. Extant entity recognition methods used for more structured text are not sufficient because message posts present additional challenges: the posts contain many typographical errors, larger variety of possible names, terms and abbreviations specific to Internet posts or a particular message board, and mentions of the authors' personal lives. The main contribution of this paper is a system to de-identify the authors of message board posts automatically, taking into account the aforementioned challenges. We demonstrate our system on two different message board corpora, one on breast cancer and another on arthritis. We show that our approach significantly outperforms other publicly available named entity recognition and de-identification systems, which have been tuned for more structured text like operative reports, pathology reports, discharge summaries, or newswire.

BMC Bioinformatics, 12(3):73

2006

Shawndra Hill, Deepak Agarwal, Robert Bell, Chris Volinsky

Building an Effective Representation for Dynamic NetworksA dynamic network is a special type of network composed of connected transactors which have repeated evolving interaction. Data on large dynamic networks such as telecommunications networks and the Internet are pervasive. However, representing dynamic networks in a manner that is conducive to efficient large-scale analysis is a challenge. In this article, we represent dynamic graphs using a data structure introduced in an earlier article. We advocate their representation because it accounts for the evolution of relationships between transactors through time, mitigates noise at the local transactor level, and allows for the removal of stale relationships. Our work improves on their heuristic arguments by formalizing the representation with three tunable parameters. In doing this, we develop a generic framework for evaluating and tuning any dynamic graph. We show that the storage saving approximations involved in the representation do not affect predictive performance, and typically improve it. We motivate our approach using a fraud detection example from the telecommunications industry, and demonstrate that we can outperform published results on the fraud detection task. In addition, we present a preliminary analysis on Web logs and e-mail networks.

Journal of Computational & Graphical Statistics, Vol. 15, No. 3, pp. 584-608 [Bibtex]

2006

Shawndra Hill, Foster Provost, Chris Volinsky

Network-based Marketing: Identifying Likely
Adopters via Consumer NetworksNetwork-based marketing refers to a collection of marketing techniques that take advantage of links between consumers to increase sales. We concentrate on the consumer networks formed using direct interactions (e.g., communications) between consumers. We survey the diverse literature on such marketing with an emphasis on the statistical methods used and the data to which these methods have been applied. We also provide a discussion of challenges and opportunities for this burgeoning research topic. Our survey highlights a gap in the literature. Because of inadequate data, prior studies have not been able to provide direct, statistical support for the hypothesis that network linkage can directly affect product/service adoption. Using a new data set that represents the adoption of a new telecommunications service, we show very strong support for the hypothesis.

Statistical Science, Vol. 21, No. 2, pp. 256-276 [Bibtex]

2005

Shawndra Hill, Abraham Bernstein, Foster Provost

Toward Intelligent Assistance for a Data Mining Process: An Ontology--Based Approach for Cost-Sensitive ClassificationA data mining (DM) process involves multiple stages. A simple, but typical, process might include preprocessing data, applying a data mining algorithm, and postprocessing the mining results. There are many possible choices for each stage, and only some combinations are valid. Because of the large space and nontrivial interactions, both novices and data mining specialists need assistance in composing and selecting DM processes. Extending notions developed for statistical expert systems we present a prototype Intelligent Discovery Assistant (IDA), which provides users with 1) systematic enumerations of valid DM processes, in order that important, potentially fruitful options are not overlooked, and 2) effective rankings of these valid processes by different criteria, to facilitate the choice of DM processes to execute. We use the prototype to show that an IDA can indeed provide useful enumerations and effective rankings in the context of simple classification processes. We discuss how an IDA could be an important tool for knowledge sharing among a team of data miners. Finally, we illustrate the claims with a demonstration of cost-sensitive classification using a more complicated process and data from the 1998 KDDCUP competition.

IEEE Transactions on Knowledge and Data Engineering, Vol. 17, Iss. 4, pp. 503-518 [Bibtex]

2003

Shawndra Hill, Foster Provost

The Myth of the Double-blind Review? Author Identification Using Only CitationsPrior studies have questioned the degree of anonymity of the double-blind review process for scholarly research articles. For example, one study based on a survey of reviewers concluded that authors often could be identified by reviewers using a combination of the author's reference list and the referee's personal background knowledge. For the KDD Cup 2003 competition's "Open Task," we examined how well various automatic matching techniques could identify authors within the competition's very large archive of research papers. This paper describes the issues surrounding author identification, how these issues motivated our study, and the results we obtained. The best method, based on discriminative self-citations, identified authors correctly 40-45% of the time. One main motivation for doubleblind review is to eliminate bias in favor of well-known authors. However, identification accuracy for authors with substantial publication history is even better (60% accuracy for the top-10% most prolific authors, 85% for authors with 100 or more prior papers).

SIGKDD Explorations, Vol. 5, Iss. 2, pp. 179-184 [Bibtex]

2012

Adrian Benton, Shawndra Hill

The Spoiler Effect?: Designing Social TV Content That Promotes Ongoing WOM Television shows are now instigating online social interactions between viewers by requesting viewers, as part of the first broadcasts, to participate in simultaneous discussions about shows. In this work, we analyze and quantify the effect of strategies used in an American reality singing show to engage viewers. We show three main results: 1) when specific tweet messages are posted on the TV screen during the show, the tweets are much more likely to be discussed by viewers than messages created by the same online social media users during the show that are not posted on the screen; 2) there is a difference between social media engagement response to specific tweet messages and more general social media content-- specific tweets with new information result in more engagement on the first show airing; and 3) there is a difference in overall response to specific social media tweet messages on the West Coast of the US since results by the time of airing have been spoiled.

CIST in Conjunction with INFORMS

2012

Shawndra Hill, Adrian Benton

Social TV: Analyzing the Impact of Social Media Strategies on Viewer Engagement “Social TV” is a term that broadly describes the online social interactions that occur between viewers while watching television. Television shows are now instigating online social interactions between viewers by requesting viewers, as part of the first broadcasts, to engage in simultaneous discussions about the shows. Examples include explicitly asking users to follow and comment on topics on the popular social media site, Twitter. A more passive approach used by TV shows displays messages from social media users on the TV screen during broadcasts. While many shows now include social media in their broadcasts, very little is known about the effectiveness of these strategies on engagement and viewership. In this work, we analyze features of the social media strategy of an American reality singing show. We show that when messages are posted on the TV screen during the show, they are much more likely to be discussed by viewers than messages created by the same online social media users during the show that are not posted on the screen, indicating that this strategy works. We also show that when a prompt is given to use a specific term in discussions online, viewers are more likely to remain engaged with the show on social media. We believe this is the first scholarly work to link real-time TV posts to realtime TV engagement response. Our contribution is a set of results, which show that prompting viewers during the show with social media posts on screen results in increased and prolonged engagement during the show.

The Second SOMA Workshop: Social Media Analytics

2012

Shawndra Hill, Aman Nalavade, Adrian Benton

Social TV: Real-Time Social Media Response to TV Advertising We link the content of real-time social media response to characteristics of the Super Bowl 2012 TV advertisements, including their advertisers' social media strategies and the categories of the products being advertised. We analyze millions of social media posts about approximately forty-five advertisements during the 2012 Super Bowl. While there are many studies that focus on the popularity of and sentiment of response to Super Bowl ads, we believe ours is the first to go deeper and analyze the textual "content" to understand better the factors contributing to specific types of social media responses among a number of TV advertising dimensions. In this study, we show that the level of online consumer engagement , measured by attracting new followers on Twitter during the Superbowl, can be linked to whether the brand that advertised had a social media strategy or not In addition, we show that sentiment response, a measure often use to quantify the effectiveness of TV ads, varies by demographic category.

ADKDD 2012: The Sixth International Workshop on Data Mining for Online Advertising and Internet Economy

2012

Shawndra Hill, Adrian Benton

Social TV: Linking Social Media Buzz to Sales

Conditionally Accepted ICIS (International Conference on Information Systems)

2011

Shawndra Hill, Benton A, Ungar L, Macskassy S, Chung A, Holmes JH

A Cluster-based method for isolating influence on TwitterThis paper demonstrates a cluster-based method to isolate influence in social networkbased observational data, where "influence" is defined to mean that one person posts about a topic online and a second person posts about the same topic because he or she read the first post. Isolating influence in observational data is difficult, because we may observe that connected people discuss the same topic in proximate periods for reasons other than influence, including homophily–connected people are similar–and exogenous shock; they may have learned of the topic from some external source. We employ a matched sample estimation technique that has been used in the past to measure influence by controlling for demographic and usage based homophily, and add to the matching scheme a cluster ID. Our contribution is two-fold: First, we provide preliminary evidence that social network-based clusters capture homophily, indicating that a network-based attribute approach may not only capture homophily but also may be used in lieu of using demographic attributes for matching similar users in scenarios when privacy preservation is a concern. Second, we show that by adding a network position attribute, a cluster ID, when matching similar users, we can isolate influence better. We believe that our approach to isolate influence can have a broad impact on problems where social networks and associated behaviors can be observed over time.

Workshop on Information Technologies and Systems

2011

Getachew Berhan, Shawndra Hill, Tsegaye Tadesse, Solomon Atnafu

Geographic Information Systems and Geo-statistics for Modeling and Mapping Endangered Species: A Case Study in Bonga Forest of Ethiopia

Global Information Technology Management Association

2010

Adrian Benton, Shawndra Hill, Lyle Ungar, Annie Chung, John Holmes

A System for De-Identifying Medical Message Board TextThere are millions of public posts to medical message boards by users seeking support and information on a wide range of medical conditions. It has been shown that these posts can be used to gain a greater understanding of patients' experiences and concerns. As investigators continue to explore large corpora of medical discussion board data for research purposes, protecting the privacy of the members of these online communities becomes an important challenge that needs to be met. Extant entity recognition methods used for more structured text are not sufficient because message posts present additional challenges: the posts contain many typographical errors, larger variety of possible names, terms and abbreviations specific to Internet posts or a particular message board, and mentions of the authors' personal lives. The main contribution of this paper is a system to de-identify the authors of discussion board posts automatically, taking into account the aforementioned challenges. We demonstrate our system on two different message board corpora, one on breast cancer and another on arthritis. We show that our approach significantly outperforms other publicly available de-identification systems, which have been tuned for more structured text like operative reports, pathology reports and discharge summaries. Our software will be available for download as open source code in the near future.

International Conference on Machine Learning Applications

2010

Getachew Berhan, Tsegaye Tadesse, Solomon Atnafu, Shawndra Hill

Knowledge Discovery from Satellite Images for Drought Monitoring: A Case Study in Ethiopia

Global Information Technology Management Association World Conference

2010

Justin Vastola, Kobi Abayomi, Shawndra Hill

Statistics for Re-Identification Signatures in Networked Data

Workshop on Information Networks

2010

Shawndra Hill, Getachew Berhan, Anita Banser, Nathan Eagle

Reality Mining AfricaCellular phones can be used as mobile sensors, continuously logging users' behavior including movement, communication and proximity to others. While it is well understood that data generated from mobile phones includes a record of phone calls, there are also more sophisticated data types, such as Bluetooth or cell tower proximity logging, which reveal movement patterns and day-to-day human interactions. We explore the possibility of using mobile phone data to compare movement and communication patterns across cultures. The goal of this proof-of-concept study is to quantify behavior in order to compare different populations. We compare our ability to predict future calling behavior and movement patterns from the cellular phone data of subjects in two distinct groups: a set of university students at MIT in the United States and the University of Nairobi in Kenya. In addition, we show how Bluetooth data may be used to estimate the diffusion of an airborne pathogen outbreak in the different populations.

AAAI Symposium on Artificial Intelligence for Development [Bibtex]

2010

Mariye Yigzaw, Shawndra Hill, Anita Banser, Lemma Lessa

Using Data Mining to Combat Infrastructure Inefficiencies: The Case of Predicting Non-payment for Ethiopian TelecomData mining and machine learning technologies for business applications have evolved over the past two decades, and are regularly applied in contemporary organizations to everything from manufacturing to online advertising in fields ranging from health care to motor racing. Unfortunately, data mining techniques are not applied as often to problems in the developing world. Despite the fact that some industries, such as banks, airlines, courts, and telecommunications firms, necessitate data storage as part of their business process. We argue that data mining could be used to reduce infrastructure inefficiencies, which is one of the largest problems faced by Africa. We demonstrate that we can potentially reduce the infrastructure inefficiency of the Ethiopian telecommunications industry by ranking customers according to their likelihood of nonpayment using a data mining approach.

AAAI Symposium on Artificial Intelligence for Development

2010

Tibebe Besheh, Shawndra Hill

Mining Road Traffic Accident Data to Improve Road Safety: The Role of Road-related Factors on Accident Severity in EthiopiaRoad traffic accidents (RTAs) are a major public health concern, resulting in an estimated 1.2 million deaths and 50 million injuries worldwide each year. In the developing world, RTAs are among the leading cause of death and injury; Ethiopia in particular experiences the highest rate of such accidents. Thus, methods to reduce accident severity are of great interest to traffic agencies and the public at large. In this work, we applied data mining technologies to link recorded road characteristics to accident severity in Ethiopia, and developed a set of rules that could be used by the Ethiopian Traffic Agency to improve safety.

AAAI Symposium on Artificial Intelligence for Development

2009

Shawndra Hill and Akash Nagle

Social Network Signatures: A Framework for Re‐Identification in Networked DataData on large dynamic social networks, such as telecommunications networks and the Internet, are pervasive. However, few methods conducive to efficient large-scale analysis exist. In this paper, we focus on the task of re-identification. Re-identification in the context of dynamic networks is a matching problem that involves comparing the behavior of networked entities across two time periods. Prior research has reported success in the domains of e-mail alias detection, author attribution, and identifying fraudulent consumers in the telecommunications industry. In this work, we address the question of "why are we able to re-identify entities on real world dynamic networks?" Our contribution is two-fold. First, we address the challenge of scale with a framework for matching that does not require pairwise comparisons to ascertain the similarity scores between networked entities. Second, we show our method is robust against missing links but less tolerant to noise. Using our framework, we provide a performance estimate for re-identification on networks based solely on their degree distribution and dynamics. This work has significant implications for re-identification problems where scale is a challenge as well as for problems where false negatives (e.g., when fraudulent consumers are not labeled as fraudulent) cannot be observed.

International Conference on Computational Aspects of Social Networks [Bibtex]

2007

Shawndra Hill, Foster Provost, Chris Volinsky

Learning and Inference in Massive Social NetworksResearchers and practitioners increasingly are gaining access to data on explicit social networks. For example, telecommunications and technology firms record data on consumer networks (via phone calls, emails, voice-over-IP, instant messaging), and social-network portal sites such as MySpace, Friendster and Facebook record consumer-generated data on social networks. Inference for fraud detection [5, 3, 8], marketing [9], and other tasks can be improved with learned models that take social networks into account and with collective inference [12], which allows inferences about nodes in the network to affect each other. However, these socialnetwork graphs can be huge, comprising millions to billions of nodes and one or two orders of magnitude more links.

5th International Workshop on Mining and Learning with Graphs [Bibtex]

2005

Shawndra Hill, Deepak Agarwal, Robert Bell, and Chris Volinsky

Tuning Representations of Dynamic Network DataA dynamic network is a special type of network which is comprised of connected transactors which have repeated evolving interaction. Data on large dynamic networks such as telecommunications networks and the Internet are pervasive. However, representing dynamic networks in a manner that is conducive to effcient large-scale analysis is a challenge. In this paper, we represent dynamic graphs using a data structure introduced by Cortes et. al. [3]. Our work improves on their heuristic arguments by formalizing the representation with three tunable parameters. In doing this, we develop a generic framework for evaluating and tuning any dynamic graph. We show that the storage saving approximations involved in the representation do not affect predictive performance, and typically improve it. We motivate our approach using a fraud detection example from the telecommunications industry, and demonstrate that we can outperform published results on the fraud detection task.

3rd International Workshop on Link Discovery

2003

Shawndra Hill

Social Network Relational Vectors for Anonymous Identity MatchingAnonymous fraudulent behavior can generate substantial financial burden and inconvenience. Moreover, the recent threat of terrorist infiltra-tion to both business and government has yielded heightened interest in anonymous iden-tity matching (AIM). Most applications of AIM require sophisticated methods robust to issues such as deliberate variation in identity attributes, missing data, and multi-source data corpora. We consider relational social network behavior, eliminating the reliance on personal identifiable data for identity matching. In particular, we consider problems that can be characterized by personal communication networks. We evaluate a proposed implementation of a social network vector-space relational model for AIM on Cite-Seer, a research publication citation database.

International Joint Conference on Artificial Intelligence

2002

Abraham Bernstein, Scott Clearwater, Shawndra Hill, Claudia Perlich, Foster Provost

Discovering Knowledge from Relational Data Extracted from Business NewsThousands of business news stories (including press releases, earnings reports, general business news, etc.) are released each day. Recently, information technology advances have partially automated the processing of documents, reducing the amount of text that must be read. Current techniques (e.g., text classification and information extraction) for full-text analysis for the most part are limited to discovering information that can be found in single documents. Often, however, important information does not reside in a single document, but in the relationships between information distributed over multiple documents. This paper reports on an investigation into whether knowledge can be discovered automatically from relational data extracted from large corpora of business news stories. We use a combination of information extraction, network analysis, and statistical techniques. We show that relationally interlinked patterns distributed over multiple documents can indeed be extracted, and (specifically) that knowledge about companies' interrelationships can be discovered. We evaluate the extracted relationships in several ways: we give a broad visualization of related companies, showing intuitive industry clusters; we use network analysis to ask who are the central players, and finally, we show that the extracted interrelationships can be used for important tasks, such as for classifying companies by industry membership.

8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Bibtex]