Keywords: text analysis, graph theory, spectral embedding, clustering
In 2012, Carey Priebe et al. defined the concept of "priviness" in a co-authorship network. In this formulation, each author is "privy" (i.e. 0-privy) to topics on which the author has published, and author A is b-privy to any topic T if the graph distance between A and the set of authors privy to T is b. Based on the assumption that interesting things happen when disparate topics are combined, the 2012 paper developed a log-odds ratio statistic based on priviness to test the hypothesis that one group of authors is more/less likely to combine knowledge in two topics together than another group of authors. In this talk I will take a slightly different approach to priviness, in which I use the topics to augment the co-authorship network, then use adjacency spectral embedding, followed by inference, to find topic relevant structure in the subsequent embedding. This will be illustrated on a set of 90K abstracts from arXiv.