A0492
Title: Change detection in dynamic networks using flexible multivariate control charts
Authors: Jonathan Flossdorf - TU Dortmund University (Germany) [presenting]
Carsten Jentsch - TU Dortmund University (Germany)
Roland Fried - TU Dortmund University (Germany)
Abstract: The focus is on the identification of differences in dynamic networks, i.e. in a sequence of networks between various time points. This task is important for statistical procedures like two-sample tests or change-point detection. Due to the rather complex nature of dynamic network data, the complexity is typically reduced to a metric or some sort of a model based on these metrics. However, the reduction in network metrics can result in a heavy information loss. Hence, understanding their behaviour in various change scenarios is crucial. We present a categorization of different types of changes that can occur in dynamic network data. We analyze the suitability and limitations of common network metrics in such situations with respect to their mathematical properties and give comprehensive explanations of their behaviour. This leads to well-founded advice on which metrics to use in various application scenarios. Based on this foundation, we develop an online monitoring approach usable for flexible network structures and types of changes. It uses a sound choice of a set of the analyzed network metrics that are jointly monitored in a suitable multivariate control chart scheme, which performs superior to univariate analysis and enables both parametric and non-parametric usage. All our findings are supported by extensive simulation studies and real-world examples.
