Research On Crucial Technologies Of Email Communication Network Change Detection

The wide use of email communication made it an important way to discover the relationships among people in social organization by email mining. The research of the dynamic of email communication network is of great significance for analyzing real organizational behaviors and warning organizational events. Dynamic modeling network is lack of sensitive awareness for practical network anomaly, and the high computation complexity or the application limit made existing network change detection approaches less useful. In this paper, some dynamic network analysis technologies are modified and integrated for constructing the framework of email communication network change detection. Research topics, such as longitudinal network creation, fast network stability discrimination, and network changing behaviors detection are studied intensively, including crucial aspects of:1. Longitudinal network based on exponential smoothing overlap is created. In order to reflect the changes of communication network, a series of different period network were connected sequentially to form the longitudinal network. There are unavoidable uncertainties when short time email data are used to reflect the real social relationships. So the exponential smoothing method was chosen to correct the uncertainties with the consideration of history data, and the selection of parameters in the correction process was analyzed.2. A method of fast stability discrimination in communication network based on networks'similarity is proposed. By a local sensitive hash algorithm, high dimensional features set of network entities was mapped into a bit string which is used to evaluate the difference between networks quickly. The developing process in the past of communication network was discovered by clustering the communication network of near time sequence, and the stable period of network can be found. The effectiveness of this approach is testified when experimenting on the simulated network and real network.3. A method to detect the change of various network features based on statistical analysis is proposed. Technologies of multivariate nonparametric statistical process control were introduced to break the limit in the requirement of particular data distribution. With the introduced approach, the statistic change of different network features can be detected by supervising multi measuremens of network concurrently. And the specific trend, degree and time of the change in network can be informed. When test on Enron email data set, the detection results conform with the real organizational behaviors.