Vertex Nomination via Content and Context

TL;DR

This paper introduces a vertex nomination method combining content and context in attributed communication graphs to identify interesting or suspicious actors, demonstrated on the Enron email dataset.

Contribution

It develops a joint model leveraging both communication content and network structure, improving vertex nomination accuracy over models using only one source.

Findings

Joint content and context modeling enhances nomination performance.

Experimental results on Enron data validate the approach.

Content-only or context-only models perform worse than combined models.

Abstract

If I know of a few persons of interest, how can a combination of human language technology and graph theory help me find other people similarly interesting? If I know of a few people committing a crime, how can I determine their co-conspirators? Given a set of actors deemed interesting, we seek other actors who are similarly interesting. We use a collection of communications encoded as an attributed graph, where vertices represents actors and edges connect pairs of actors that communicate. Attached to each edge is the set of documents wherein that pair of actors communicate, providing content in context - the communication topic in the context of who communicates with whom. In these documents, our identified interesting actors communicate amongst each other and with other actors whose interestingness is unknown. Our objective is to nominate the most likely interesting vertex from all vertices with unknown interestingness. As an illustrative example, the Enron email corpus consists of communications between actors, some of which are allegedly committing fraud. Some of their fraudulent activity is captured in emails, along with many innocuous emails (both between the fraudsters and between the other employees of Enron); we are given the identities of a few fraudster vertices and asked to nominate other vertices in the graph as likely representing other actors committing fraud. Foundational theory and initial experimental results indicate that approaching this task with a joint model of content and context improves the performance (as measured by standard information retrieval measures) over either content or context alone.