Monitoring Social Networks in Phase I using Zero-Inflated Poisson Regression Model

In this paper, zero-inflated Poisson (ZIP) regression was assumed as an underlying model to generate network data. This model can be an appropriate model if the network data is sparse and produced with two processes, one generates only zeros and the other generates count data that follow the Poisson model, the two parameters of the model are functions of variables here referred to as similarity variables. The performance of the Likelihood Ratio Test (LRT), a Combined Residual-Square Residual (R-SR), and Hotelling's T^2 control charts was investigated in networks based on the ZIP regression model in Phase I. Traditionally, in Phase I the parameters of the model are unknown and need to be estimated. One needs to be sure the process is stable and the changes are detected and removed. The performance of our proposed methods is compared using simulation when parameters slope and intercept are under step changes. Signal probability was recorded as a comparison measure. The simulation results show that the LRT outperforms two other methods significantly in terms of signal probability. The efficiency of methods was also examined in real Enron data set