Improved Change Detection in Longitudinal Social Network Measures Subject to Pattern-of-Life Variations

This paper describes the challenges posed by pattern-of-life variations when carrying out automated detection of abnormal events (change detection) in longitudinal (over-time) social network data sets using standard social network measures. In this paper we present a new scheme for substantially removing pattern-of-life variations from longitudinal social network measures. This new approach is based on a model in which pattern-of-life variations are modeled as time-dependent periodic multiplicative weights on the likelihood of initiating a new post in a social network. Unfortunately, analysis of real-world social network data reveals that the time-dependent weights change over time as well. Therefore, an approach for adaptively determining the time-dependent periodic multiplicative weights has been developed. A complete methodology for Adaptive Multiplicative Compensation for Pattern-of-Life variations is described and the methodology is tested on a suitable social media data set. The impact of pattern-of-life variations on the test over-time data set is reduced by up to a factor of 4X by the algorithm presented. The impact on the occurrences of false positive events (labeling a time point as a “change” when it is not) and the impact on the occurrences of false negative events (labeling a time point as “normal” when it really represented a change) clear in the test data set.