Effect of Aggregation when Monitoring Social Networks

The goal of monitoring social networks is to detect anomalies in a network as quickly as possible. Detection of such anomalies can be used to identify suspicious individuals or groups of individuals. In social network analysis aggregation over time is ubiquitous. Though this may be necessary for practical reasons, whenever data is aggregated (from timestamp to count in equal units of time) there is a loss of information. One example of such temporal aggregation is Sparks and Paris (2017) who collected tweet informatics in real time and aggregate the data to an hourly level.This so-called temporal aggregation is common practice in many research areas including monitoring. Zwetsloot and Woodall (2018) found that currently no detailed guidelines exist for selecting appropriate levels of temporal aggregation, also little research exists which studies the effect of temporal aggregation on the performance of monitoring tools.Therefore, in this project we focus on the use of temporal aggregation within the field of statistical process monitoring (SPM). This project will explore the effect of temporal aggregation with a focus on network data from social networks. The first objective of this project is to study the effect of aggregation on monitoring networks. I will illustrate the effect of temporal aggregation using well known case studies and simulation. In the second part guidelines and strategies are presented to select appropriate levels of temporal aggregation. For this new monitoring models will be developed.Schuh et al. (2013) showed that when monitoring univariate count data the detection speed increases as there is less temporal aggregation. They advise using time-between events models. Based on this result we hypothesize that lower levels of aggregation will increase the inspection speed for network monitoring as well. The third part of this project will explore how build a model for real time monitoring of network data based on the concept of time-between events models.