A two-stage model for change point detection in large-scale weighted directed social networks using a MEWMA chart

Due to political, cultural, and social reasons, monitoring social network has increasingly attracted the attention of both practitioners and scholars. Most studies in this area suffer from five main shortcomings: (1) the use of case-based attributes to examine the network behavior which limits their applicability to special networks, (2) neglecting the estimation of real-time of change in network behavior, (3) lack of control charting methods for monitoring weighted-directed networks, (4) ignoring the correlation structure among attributes under investigation, and (5) their limited applicability to only small-scale networks. This study proposes a two-stage method to remedy all five mentioned shortcomings for monitoring large-scale weighted-directed social networks. Using general attributes for detecting network disturbances enables the proposed method to be employed in a wide range of social network applications in contrast to existing studies that have focused on special networks. Since estimating the real time of change in network behavior called “change point” reduces the required time and cost for recognizing the cause of deviation, the proposed method focuses on both detecting network disturbances and discovering the change point by taking the correlation structure among network attributes. To do this, a topology control algorithm based on the connected dominating set is used the first stage to shrink the network. Then, in the second stage, the likelihood ratio test (LRT) is used for change point estimation in the shrunk network obtained in the first stage. Afterward, using two numerical examples, the capability of the proposed two-stage methodology is evaluated and compared with the competing monitoring schemes. The obtained results represent that the developed two-stage methodology works better than the competing methods in terms of accuracy and precision criteria.