Publications and Papers
Radford and Danneman (2019). Using Attention for Improved interpretation of Citation Network Models
In this paper, Ben Radford and I demonstrate interpretable link prediction and node classification on the citation network of social science articles from the Semantic Scholar database. Links, or citations, are modeled as a function of the titles and abstracts of article pairs. Because the link prediction portion of this model is inductive, links between papers can be predicted even for those papers that are unknown to the model at training time. Additionally, we model the journal venue for each paper given the paper’s title and abstract. While most neural network models are “black boxes” in that their complicated functional forms resist intuitive interpretations, we show that the addition of attention layers can be informative for researchers interested in understanding the predictors on network models of scholarly research products. Full paper (pdf).
Danneman and Esarey (2015). A quantitative method for substantive robustness assessment. Political Science Research and Methods.
Empirical political science is not simply about reporting evidence; it is also about coming to conclusions on the basis of that evidence and acting on those conclusinos. But whether a results is substantively significant — strong and certain enough to justify acting upon the belief that the null hypothesis is false — is difficult to objectively pin down, in part because different researchers have different standards for interpreting evidence. Instead, we advocate juding results according to their “substantive robustness,” the degree to which a community with heterogeneous standards for interpreting evidence would agree that the result is substantively significant. We ilustrate how this can be done using Bayesian statistical decision techniques. Judging results in this wey yields a tangible benefit: false positive are reduced without decreasing the power of the test, decreasing the error rate in published results. Full text (pdf).
Danneman and Ritter (2014). Contagious rebellion and preemptive repression. Journal of Conflict Resolution.
Civil conflict appears to be contagious—scholars have shown that civil wars in a state’s neighborhood make citizens more likely to rebel at home. However, war occurs when both rebels and the state engage in conflict. How do state authorities respond to the potential for civil conflict to spread? We argue that elites will anticipate the incentive-altering effects of civil wars abroad and increase repression at home to preempt potential rebellion. Using a Bayesian hierarchical model and spatially weighted conflict measures, we find robust evidence that a state will engage in higher levels of human rights violations as civil war becomes more prevalent in its geographic proximity. We thus find evidence that states violate rights as a function of the internal politics of other states. Further, we argue authorities will act not to mimic their neighbors but rather to avoid their fate. Full text (pdf).
Danneman and Beardsley (2015). International mediation and the problem of insincere bargaining. Emerging Trends in the Behavioral and Social Sciences.
Much of the existing literature on international mediation has focused on its capacity to assuage various barriers to efficient bargaining. Often implicit in these studies is the assumption that the parties involved are negotiating sincerely, in good faith. Less studied are the incentives of the principals to use negotiation as a stalling tactic: while the talks are underway, one or more of the parties takes steps to improve its armed-conflict capabilities and/or resolve. This paper argues that mediators can use leverage—threats or uses of retaliation and punishment—to enhance the possibility that the principals negotiate in good faith. Evidence supportive of this argument is explored empirically using cross-national data on intrastate disputes. Third parties that are geographically proximate and major powers—characteristics of third parties that are both willing and capable to enforce good-faith bargaining —greatly improve the likelihood that a mediation initiative will achieve an agreement. Full text (pdf).
Danneman and Heimann (2014). Social Media Mining with R. Packt Publishing.
The growth of social media over the last decade has revolutionized the way individuals interact and industries conduct business. Individuals produce data at an unprecedented rate by interacting, sharing, and consuming content through social media. However, analyzing this ever-growing pile of data is quite tricky and, if done erroneously, could lead to wrong inferences.
By using this essential guide, you will gain hands-on experience with generating insights from social media data. This book provides detailed instructions on how to obtain, process, and analyze a variety of socially-generated data while providing a theoretical background to help you accurately interpret your findings. You will be shown R code and examples of data that can be used as a springboard as you get the chance to undertake your own analyses of business, social, or political data.
The book begins by introducing you to the topic of social media data, including its sources and properties. It then explains the basics of R programming in a straightforward, unassuming way. Thereafter, you will be made aware of the inferential dangers associated with social media data and how to avoid them, before describing and implementing a suite of social media mining techniques.
Social Media Mining in R provides a light theoretical background, comprehensive instruction, and state-of-the-art techniques, and by reading this book, you will be well equipped to embark on your own analyses of social media data.
Danneman (2013). Committing to Bargain: How Mediation Contributes to the Onset and Success of Peace Talks. PhD Dissertation, Emory University Department of Political Science.
Feed Forward and Convolutional Nets: An Intro (May, 2020)
I got to interact with the wonderful folks at
RuleBreaker: Categorical Correlations as Probabilistic Rules (June, 2018)
I had a great time at the BSides Asheville infosec conference. Lots a great speakers on a wide range of applied cyber security topics. I spoke about using association rule mining to identify strong categorical correlations in big cyber data. You can then use these “rules” to identify discrepancies in, e.g. netflow data. If you prune the rules a bit, and choose the features carefully, the breakages of these rules end up being quite interesting. The slides are here, and you can catch the video here (see the 3:11:00 mark).
Anomaly Detection: Algorithms, Trade Offs, and Best Practices (May, 2018)
I had a really nice time speaking at the Data Science, Machine Learning, and AI Meetup yesterday, hosted by Syntelli in the Charlotte area. I gave a talk the skimmed the literature on anomaly detection in IID data, and gave a preview of some comparative evidence being built by a research group I am part of at DARPA. The slides are available, and the code will be soon.
K-Means Spill Trees at Scale: Catching Credential Misuse and Domain Squatting (October, 2017)
I had the pleasure of speaking at the first (annual?) Conference for Applied Machine Learning in Information Security (CAMLIS) this past October. Big thanks to Keegan Hines for doing a ton of the heavy lifting in organizing, and to Data Machines Corp. for co-sponsoring. My talk covered how I was frustrated by the lack of scalability and sensitivity to parameter choices I experienced when using various space-partitioning algorithms as pre-processing tools for nearest-neighbor models. As such, I took a crack at writing a spill tree with K-Means underlying the splitting, and soft clustering that allayed some of the issues with parameterization. Here are the slides; code is forthcoming, pending client agreement.
Social Media Mining (day-long course; January, 2015)
I got to deliver part of a day-long course on social media mining, following the release of Social Media Mining with R. Materials available upon request.
Unsupervised Sentiment Scaling
Big thanks to the folks at NLP-DC for having me give a talk on unsupervised sentiment scaling. At a high level, the idea is to use a bad-of-words approach, and take each word as a marker of how much an authors “supports” and underlying idea. Then, you can use Item Response Theory (or any other matrix scaling technique) to extract the assumed, underlying continuum of sentiment. This method is described in depth in Social Media Mining with R.
Geospatial Anomaly Detection
Big thanks to George Mason University’s Department of Computational and Data Sciences for having me out to give a talk on geospatial anomaly detection. The talk centered on the use of treating unsupervised learning as a supervised problem, and on robust, regression-based ways to model space and spatial correlates. I also gave a variant of this talk at a Data Science DC Meetup. One version of the slides is here.