Community-Attribute Models for Bibliographic Reference Information via Dynamic Graph Evolution

Abstract

We present CAMBRIDGE, the first technique for evolutionary analysis of dynamic graphs via a community-attribute graph model. Community-attribute models have been shown to be superior to models conventionally used for evolutionary analysis, particularly in modeling intercommunity structures in networks where communities exhibit dense overlaps. Thus, our use of a community-attribute model for analysis of a bibliographic network evolving in time allows us to observe not only the evolution of discrete clusters, but also the evolution of the ‘core’ of nodes which are strongly linked to multiple communities simultaneously. In particular, our approach allows us to observe and quantify how the sibling communities resulting from community-splitting events share and compete for external intercommunity influence inherited from parent communities. We present evidence that indicates that in such splitting events, highly connected nodes which were part of the parent networks ‘strong intercommunity ties’ become concentrated in the siblings’ intersection, whereas highly-connected nodes that are part of ‘weak intercommunity ties’ are dispersed to the individual sibling communities. We discuss the implications of our findings for the field of evolutionary graph analysis, and address the evident promise of dynamic community-attribute models in providing fully generative models for dynamic networks.