Change point detection in relational event history data based on moving window approach with optimal window length.

Abstract

This paper proposes a strategy to study the social dynamics in timestamped data. The analysis is based on the Relational Event History (REH). It introduces Apollo 13 data to illustrate the study objective. The goal of this study is to detect change points in relational event history data based on moving window approach with optimal window length. One of the methods used in the study is the Relation Event Model. The Relational Event model is a method to study social networks over time. This method is later on complemented with the more dynamic method called the moving window approach. The results of the moving window approach, expressed in parameters, are the input variables for the next method called change point detection. Change point detection is introduced to study the change in parameters over the different windows, to analyze the social interaction patterns that can be found in the Apollo 13 data. Data cleaning was performed on the Apollo 13 dataset to make it suitable for the analysis. The analysis is performed in R. Seven effects are proposed to fit the Relational Event Model. Five moving window lengths are considered in order to find the optimal window size. These window sizes are complimented by three different overlapping percentages. Change points are detected using the ‘MedoidAI’ application. ‘MedoidAI’ is an R/Shiny app for time-series segmentation and changepoint detection tasks. In the application two different algorithms are considered to detect multiple change points. These algorithms are the Binary Segmentation and the PELT algorithm. Both are considered and compared in their ability to detect change points. The results of the study show an optimal window length at 0.5 hours with 50 percent overlap. This window length showed the lowest BIC score. The optimal window can detect change points that are more specific than the results seen at a larger window size. While the smaller window size showed a results that seems more sensitive to the datapoints over time. This implies that the optimal window could be the optimal size to detect change points in the Apollo 13 dataset.