Online monitoring of dynamic networks using flexible multivariate control charts

Abstract

AbstractChange-point detection in dynamic networks is a challenging task which is particularly due to the complex nature of temporal graphs. Existing approaches are based on the extraction of a network’s information by the reduction to a model or to a single metric. Whereas the former one requires restrictive assumptions and has limited applicability for real-world social networks, the latter one may suffer from a huge information loss. We demonstrate that an extension to a well-balanced multivariate approach that uses multiple metrics jointly to cover the relevant network information can overcome both issues, since it is applicable to arbitrary network shapes and promises to strongly mitigate the information loss. In this context, we give guidelines on the crucial questions of how to properly choose a suitable multivariate metric set together with the choice of a meaningful parametric or nonparametric control chart and show that an improper application may easily lead to unsatisfying results. Furthermore, we identify a solution that achieves reasonable performances in flexible circumstances in order to give a reliably applicable approach for various types of social networks and application fields. Our findings are supported by the use of extensive simulation studies, and its applicability is demonstrated on two real-world data sets from economics and social sciences.