Project Description: |
Communication networks can be modeled as dynamic graphs with time-varying
edges. Real-life events may cause communications that are unusual in
either volume or pattern. Event detection focuses on identifying these
unusual patterns in communication graphs. It is crucial for counter
terrorism, network surveillance and traffic management. Most event
detection methods only focus on network-wide events. However, local events,
that is those associated with only a few individuals, are more common
and of significant interest. In this study, we develop an approach to
detect those events with only local impacts. The difficulty of this task
is that the networks usually consist of heterogeneous vertices
representing people who play different roles in the community or
organization. These vertices have different communication patterns and
thus, different definitions of "anomaly". Methods treating them without
discrimination would give very noisy results and are usually biased by a
specific kind of vertices. In our approach, we define a set of metrics to
characterize people's communications from different viewpoints, and
cluster them into groups with similar behaviours. Deviations from previous
behaviours and typical cluster behaviours are regarded as evidence of
events or anomalies. This approach will be tested on both synthetic data
and real email data such as the Enron dataset.
|