Median graphs and anomalous change detection in communication networks

Abstract

To improve the network management of large enterprise data networks we propose a novel approach to abnormal network change detection. Periodic observations of logical communications within a network are represented as a time series of graphs. Based on two different graph distance measures, the concept of median graph is introduced. The median of a time series of graphs is a graph that is most representative of all graphs in the series. In the application to time series of communication networks considered in this paper, constraints exist that greatly decrease the computational complexity of the construction of median graphs. An application of median graphs to the analysis of a large data network is given.