I am broadly trained, with most of my work focusing on interpretable strategies for analyzing text, network, and event data.

• 🔭 I’m currently working on developing the package for a bespoke Bayesian IRT Model that returns identifiable dimensions.
• 👯 I’m always looking for new, interesting, projects. Reach out via LinkedIn or email.
• ⚡ Fun fact: in a previous life, I was a qualitative threat hunter specializing in online extremist communities.
• 😄 Pronouns: she/her

My current main line of research with a lab at Duke is on applying a novel Item Response Theory model--- called IRT-M---that produces interpretable underlying dimensions. The package is available at the IRT-M Github repository and can be installed for R using your preferred GitHub-supporting package installer, such as install_github() call from devtools or pgk_install() from the pak library.

Our current projects use IRT-M to:

• Identify latent "threat" dimensions for perception of threats to cultural, economic, health, religion, and society along with support for immigration and the European Union in the Februrary-March 2021 Eurobarometer survey (round 94.3).
• Extract a latent "satisfaction with the status quo" dimension from more than 100,000 respondents to rounds 4, 5, and 6 of the Afrobarometer survey in 34 countries. We correlate the satisfaction dimension with the local occurrance of instability and anti-state mobilization.

I have a long-running research program on using machine learning and event data to estimate the behavior of actors who contribute to geopolitical threats. This is a tricky problem technically because data on internal processes is hard (or impossible!) to access. However, having estimates of change is important for improving risk evaluations because such actors decrease security and destabilize governance and trade. Projects in that line of research include:

• Subject to Change, which uses news reporting to estimate change years for civil war conflict actors. The peer-reviewed paper is Accepted at Political Science Research and Methods (PSRM). You can explore the data by leaving GitHub and loading the interactive dashboard. Heads up: it will take about a minute to load.

• In addition to the GitHub repository, you can check out a presentation of the project, via slides I prepared for a 2024 talk.

• Growth Trap models the on-the-ground implications of internal organizational change in an opaque, high-impact, geopolitical actor. It models changes in target preferences in al-Qaeda in the Arabian Peninsula (AQAP) before and after the group expanded locally. The takeaway argument from the research is that while AQAP became stronger in the sense that they had/have access to more resources, they also became less of a global threat actor, likely due to the hyperlocal preferences of their new recruitment base.

I have created synthetic data needed for custom applications.

• Negotiations Sim contains code to create synthetic text negotiation data with two competing frames of interest that may or may not overlap. I created it to help me develop and test methods to model and predict political outcomes in the data, including gridlock. The repository includes classification performance benchmarks at different levels of ground-truth labels for five popular machine learning algorithms (K-Nearest Neighbors, Naive Bayes, Random Forest, Ridge Regression, and Support Vector Machines). The tag levels (5%, 10%, or 25%) are intended to provide a guideline for researchers looking for a guideline for how much of this type of data to tag. Across all levels of ground truth, Random Forest performed best.

• IRT-M Synth contains a pipeline to create a synthetic sample of real survey data. The code is specifically tuned for the IRT-M package vignette, but is adaptable for other forms of tabular data. The data is injested, cleaned, and sampled in Python (via Jupyter Lab notebooks), the sample is simulated using YData's ctgan generative adversarial network model (via a Google Colab notebook and GPU). The output is verified by using it in an IRT-M application.

I have designed and implemented custom code to create quantitative measurements applied to qualitative contexts.

My work here focuses on techniques for (1) network and (2) unstructured text data.

• Dominos, a model that simulates network-based recruitment and collapse of a community. The model tests the claim that users recruited into a community together tend to also leave together if any of them become unhappy. The simulation asks what network structures are most resilient to these disruptions. You can check out a design mockup of the project here and look at some of the results in the form of a presentation that I gave to a working group on Organizational Studies organized by Stanford's Center for Advanced Study in the Behavioral Science.

The biggest takeaway is that "small world" type networks, where users have lots of local connections are the most stable when people start to leave. Conversely, communities that start with a few highly connected influencers are vulnerable to collapse... even if the influencers were part of the group before the new people came in!

• Simulated Degradation, takes an empirical network and a desired linear model and determines how robust the model results are to missing network data. The insights are twofold. First, looking at model instability helps understand which ties are important to the model results. These are the ties to focus on in the future. Second, this analysis helps verify that the model reflects the underlying network process.

I have given workshops on applying data science concepts to real-world data and problems. One such workshop for the United Nations Department of Peacekeeping (UNDPO) presented AI and NLP for peacekeeping data.

A deep dive of ML techniques used to profile the evolution of one such actor can be seen in a 2022 presentation that I gave to the Sciences Po médialab.

In 2024, I presented on using NLP to estimate threat actor changes from recorded behavior. The benefit of this technique is that it is interpretable and can be done at scale.