• View
• Edit
• History
• Print

Characterizing Structure In Anomalous Observations

Correlations between anomalous activity patterns of actors in an interaction network can yield pertinent information about complex social processes: a significant deviation from normal behavior, exhibited simultaneously by multiple actors, can provide evidence for some underlying event involving those actors -- i.e., a multilateral event. We introduce a new nonparametric Bayesian latent variable model that explicitly captures correlations between actors' anomalous behavior and uses these shared deviations from normal activity patterns to identify and characterize multilateral events. In our model's generative story, a partition of actors in an interaction network is drawn from a Chinese Restaurant Process. The counts for observed interactions between actors in the network are then drawn from a Gamma-Poisson distribution which is parametrized in part by an actor-specific background rate and in part by a group-specific deviation shared by all actors in the group. We jointly infer background rates, deviation factors, and the latent groups of actors which capture correlations in extreme interaction counts between actors.

We apply our model to a new international relations data set -- Global Database of Events, Language, and Tone (GDELT) -- which consists of interaction data of the form "who did what to whom" automatically extracted from news articles. We infer latent groups of country relations that receive anomalous amounts of press coverage at the same time and correspond to real-world multilateral events.