Group Detection and Link Analysis

Jeremy Kubica

Abstract

Discovering underlying structure from co-occurrence data is an important task in a variety of fields, including: insurance, intelligence, criminal investigation, epidemiology, biology, human resources, and marketing. We attempt to find one type of underlying structure, groupings of entities, from the co-occurrence data. To this end, we propose a probabilistic model of co-occurrence generation from the underlying groups and an algorithm (GDA) for finding these groupings. This approach combines observational co-occurrence data with entities' background demographic information, allowing us to utilize both types of data. The parameters of the model are learned via a maximum likelihood search. We show examples of group detection on several real-world and artificial data sets. We also introduce k-groups, a second algorithm that makes group detection tractable on large data sets.

Back to the Main Page