Classification and Link Prediction in a Text-Mined Hypergraph of Academic Relationships

• In Event: 472 - Section on Methodology Paper Session. Innovative Methodological Applications to Data Analysis

Abstract

Nearly all social and professional relationships exhibit multiplexity, consisting of many different kinds of connections. Using a set of faculty CVs from a single academic institution, we represent academics as linked in a complex hypergraph of associations, each inferred from the extraction of co-occurring named entities (e.g., places and organizations) in CVs. This results in an edge-sparse, link-dense hyper-graph with weighted connections. This structure is first verified by showing that academics with similar edge-incidence vectors--that are close in our hyper-graph--tend to be in the same department. We then use a novel model to generate dyadic predictions based on both the mass of a link (sets of weighted edges) and the distributional similarity of links through shared neighbors. Such prediction tasks have been well-explored for unipartite social networks, but in hyper-graphs where edge-types vastly out-number nodes, accounting for edge-similarity is crucial. The model is tested by predicting lesioned parts of the existing hyper-graph. Our analysis recovers a majority of the lesioned link mass and consistently outperforms the use of link mass alone. Results support the text-mined, named entity hyper-graph of multi-faceted relationships and highlight the importance of link mass and link similarity in driving social relations.

Authors

• Aaron Michael Gerow, University of Chicago

• Bowen Lou, University of Chicago

• Eamon Duede, University of Chicago

• James A. Evans, University of Chicago