This project focused on the development of analytical tools capable of
capturing the diversity of science. Our primary goal was to develop
and implement new methods, grounded in the computer science literature
(specifically statistical topic modeling and social network analysis),
for analyzing the impact of policy interventions on the diversity of
science. Our tools moved beyond traditional "citation-counting"
methods that focus only on the rate of scientific innovation.

We developed several new statistical topic modeling techniques for
analyzing scientific publications and patents. These include an
experimental framework for identify unlabeled similar document pairs,
new methods for quantifying and improving the quality or "coherence"
of topics produced by topic models, an investigation of correlations
and anti-correlations between inferred topics, a method for
pre-analysis phrase xdiscovery, and a method for post-analysis
discovery and presentation of topic-specific phrases.

We also focused on jointly modeling three aspects of scientific and
commercial communities: structure, content, and dynamics. We developed
new methods for analyzing the dynamics of information transfer between
entities, in order to quantify the ways in which these changes are
driven by "social" factors. Our methods draw on ideas from Bayesian
tensor factorization (shown to be effective for community detection in
social networks and topic discovery in text corpora) to form a unified
framework for modeling the structure, content, and dynamics of
communities. To complement this work, we also developed a new
general-purpose Bayesian admixture model for discovering and
visualizing topic-specific subnetworks in collaboration networks.

Finally, we explored issues surrounding fairness, accountability,
transparency, and ethics that arise when studying social processes,
such as those that underlie scientific publication and patenting.

Last Modified: 01/04/2017
Modified by: Hanna M Wallach