Active discovery of network roles for predicting the classes of network nodes

摘要: Nodes in real world networks often have class labels, or underlying attributes, that are related to the way which they connect other nodes. Sometimes this relationship is simple, for instance nodes of same may be more likely connected. In cases, however, not true, and link a network exhibits different, complex their attributes. Here, we consider know how connected, but do labels relate links. We wish identify best subset label order learn between node attributes can then use discovered accurately predict rest We present model identifies groups with similar patterns, call roles, using generative blockmodel. The predicts by learning mapping from roles maximum margin classifier. choose according an iterative margin-based active strategy. By integrating discovery classifier optimisation, process adapt better represent classification. demonstrate exploring selection networks, including marine food web English words. show that, contrast classifiers, achieves good classification accuracy range different relationships