Building networks of shared research interests by embedding words into a representation space

Departments within a university are not only administrative units, but also an effort to gather investigators around common fields of academic study. A pervasive challenge is connecting members with shared research interests both within and between departments. Here I describe a workflow that adapts methods from natural language processing to generate a network connecting $n=79$ members of a university department, or multiple departments within a faculty ($n=278$), based on common topics in their research publications. After extracting and processing terms from $n=16,901$ abstracts in the PubMed database, the co-occurrence of terms is encoded in a sparse document-term matrix. Based on the angular distances between the presence-absence vectors for every pair of terms, I use the uniform manifold approximation and projection (UMAP) method to embed the terms into a representational space such that terms that tend to appear in the same documents are closer together. Each author's corpus defines a probability distribution over terms in this space. Using the Wasserstein distance to quantify the similarity between these distributions, I generate a distance matrix among authors that can be analyzed and visualized as a graph. I demonstrate that this nonparametric method produces clusters with distinct themes that are consistent with some academic divisions, while identifying untapped connections among members. A documented workflow comprising Python and R scripts is available under the MIT license at https://github.com/PoonLab/tragula.