Efficient and Time Scale-Invariant Detection of Correlated Activity in Communication Networks

In many real-world networks, interactions between entities are observed at specific moments in continuous time, such as email, SMS messaging, and IP traffic. The majority of methods for analyzing such data first aggregate communication over designated time blocks, resulting in one or more discrete time series, to which existing tools can be applied. However, regardless of how the block lengths are chosen, discretizing time inherently introduces information loss and biases analysis towards patterns occurring at the designated time scale, effects which can be especially pronounced in networks with a high degree of temporal variability. Due to this, there has been increasing interest in using stochastic point processes to model network activity. We present a novel approach based on such models to detect times and sets of entities with temporally correlated recent activity. We develop efficient algorithms and compare our approach to existing and baseline methods through experiments on synthetic and real-world data.