Workshop in Applied Statistics presentation by Tina Eliassi-Rad, Northeastern University.

Abstract: Given a network, how can we automatically discover roles (or functions) of nodes? Roles compactly represent structural behaviors of nodes and generalize across various networks. Examples of roles include "clique-members," "periphery-nodes," "bridges," etc. Are there good features that we can extract for nodes that indicate role-membership? How are roles diffevent from communities and from equivalences (from sociology)? What are the applications in which these discovered roles can be effectively used? In this talk, we address these questions, provide unsupervised and supervised algorithms for role discovery, and discuss why roles are so effective in many applications from transfer learning to re-identification to anomaly detection to mining time-evolving networks and multi-relational graphs.