Nearly all personal relationships exhibit a multiplexity where people relate to one another in many different ways. Using a set of faculty CVs from multiple research institutions, we mined a hypergraph of researchers connected by co-occurring named entities (people, places and organizations). This results in an edge-sparse, link-dense structure with weighted connections that accurately encodes faculty department structure. We introduce a novel model that generates dyadic proposals of how well two nodes should be connected based on both the mass and distributional similarity of links through shared neighbors. Similar link prediction tasks have been primarily explored in unipartite settings, but for hypergraphs where hyper-edges out-number nodes 25-to-1, accounting for link similarity is crucial. Our model is tested by using its proposals to recover link strengths from four systematically lesioned versions of the graph. The model is also compared to other link prediction methods in a static setting. Our results show the model is able to recover a majority of link mass in various settings and that it out-performs other link prediction methods. Overall, the results support the descriptive fidelity of our text-mined, named entity hypergraph of multi-faceted relationships and underscore the importance of link similarity in analyzing link-dense multiplexitous relationships.
展开▼
