Interactive visual co-cluster analysis of bipartite graphs

Abstract

A bipartite graph models the relation between two different types of entities. It is applicable, for example, to describe persons' affiliations to different social groups or their association with subjects such as topics of interest. In these applications, it is important to understand the connectivity patterns among the entities in the bipartite graph. For the example of a bipartite relation between persons and their topics of interest, people may form groups based on their common interests, and the topics also can be grouped or categorized based on the interested audiences. Co-clustering methods can identify such connectivity patterns and find clusters within the two types of entities simultaneously. In this paper, we propose an interactive visualization design that incorporates co-clustering methods to facilitate the identification of node clusters formed by their common connections in a bipartite graph. Besides highlighting the automatically detected node clusters and the connections among them, the visual interface also provides visual cues for evaluating the homogeneity of the bipartite connections in a cluster, identifying potential outliers, and analyzing the correlation of node attributes with the cluster structure. The interactive visual interface allows users to flexibly adjust the node grouping to incorporate their prior knowledge of the domain, either by direct manipulation (i.e., splitting and merging the clusters), or by providing explicit feedback on the cluster quality, based on which the system will learn a parametrization of the co-clustering algorithm to better align with the users' notion of node similarity. To demonstrate the utility of the system, we present two example usage scenarios on real world datasets.