Abstract
Many real-world data can be represented as bipartite networks composed of two types of vertices. Paperauthor networks and event-attendee networks are the examples of bipartite networks. Detecting communities from such bipartite networks is practically important for finding similar items and for understanding the structures of the networks. In order to evaluate the goodness of detected communities from unipartite networks, Newman-Girvan modularity is often employed. For bipartite networks, Barber, Guimera, Murata and Suzuki propose bipartite modularities. This paper compares these bipartite modularities in order to understand their properties. Experimental results for synthetic bipartite networks show that (1) computation of Barber bipartite modularity is relatively fast, and (2) accuracies of communities detected by maximizing Suzuki bipartite modularity is relatively better than maximizing other bipartite modularities. In addition, we implemented a fast optimization method for detecting communities in large bipartite networks.
