Study On Statistical Quality Monitoring Methods For Network Processes

Monitoring network processes has wide applications in anomaly detection of manufacturing and supply networks,spam detection,and early diagnosis of pathological changes of celluar networks,to name a few.Research interests in network monitoring has significantly increased in recent years.A network can be characterized by its structural statistics,which are commonly used as network features for network monitoring.In many existing researches,multiple network feature statistics are monitored separately,while two important factors – the correlation among the statistics and the autocorrelation between networks at different time points are overlooked.As such,methods of monitoring network processes considering the two types of correlations are studied here.First,it is found that multiple feature statistis tend to be correlated based on analysis of a real case as well as through literature study.Monitoring multiple structural statistics of network data without taking their correlations into account may lead to a decrease of the detection power.A multivariate chart is proposed for simultaneously monitoring numbers of egdes,two-stars,and triangles for independent networks.Simulation experiments show that the proposed control chart performs better than a benchmark method when the propensity probability or the degree variability in a community has a large shift.The effectiveness of the proposed chart is illustrated by brain network data.Second,in order to explore the effect of network autocorrelation to the performance of network control chart,a simulation framework is designed for comparing some commonlyused control charts for the number of edges.Networks are simulated based on separable temporal exponential random graph models for scenrios of no autocorrelations,low,medium and high positive autocorrelations of edge counts.Results show that the increase of positive autocorrelation contributes to a substantial decrease of in-control average run length of moving-range-based control chart for the number of edges.Furthermore,it leads to a significant increase of the variation of run lengths for standard-deviation-based control chart and Poisson-based control chart.Third,with respect to autocorrelated network data,a moving window approach is adopted for collecting baseline data.Then an ARIMA model is applied for fitting network structural statistics in the window to account for the systematic variability.A standardizedresidual chart is proposed for detecting anomalous shifts of the process.The proposed method is applied to an email communication network and it successfully signals the occurrence of special events.Our studies contribute to the research of network process monitoring by considering the correlation among network structural statistics and the autocorrelation of networks over time.Moreover,the proposed methods can be widely applied in modern manufacturing and logistic processes,government and corporate management,as well as health surveillance among others.