Scientific community detection via bipartite scholar/journal graph co-clustering

This paper stems from the observation that researchers in different fields tend to publish in different journals. Such a relationship between researchers and journals is quantitatively exploited to identify scientific community clusters, by casting the community detection problem into a co-clustering problem on bipartite graphs. Such an approach has the potential of leading not only to the fine- grained detection of scholar communities based on the similarity of their research activity, but also to the clustering of scientific journals based on which are the most representative of each community. The proposed methodology is purely data-driven and completely unsupervised, and does not rely on any semantics (e.g. keywords or a-priori subjective categories). Moreover, unlike "flat" data structures (e.g. collaboration graphs or citation graphs) our bipartite graph approach blends in a joint structure both the researcher's attitude and interests (i.e., freedom to select the venue where to publish) as well as the community's recognition (i.e., acceptance of the publication on a target journal); as such may perhaps inspire further scientometric evaluation strategies. Our proposed approach is applied to the Italian research system, for two broad areas (ICT and Microbiology&Genetics), and reveals some questionable aspects and community overlaps in the current Italian scientific sectors classification.