Research On Abnormal Structure Change Discovery Technology For Dynamic Network

Dynamic network covers a wide range,computer networks,social networks,transportation networks,molecular structure networks,etc.,can use dynamic network model for correlation analysis.Dynamic network is characterized by its dynamic nature,the network structure will change with the network evolution process,these changes may exist abnormal changes that we do not want to appear.The detection of abnormal changes in dynamic networks can help us discover the network anomaly in time,and understand the trend of network development,which is of great significance in practical application.This paper takes the topology structure of dynamic network as the research object,detects the trend abnormal change and the abrupt abnormal change of dynamic network,and identifies the possible abnormal link behavior in the network.The main works of this paper is as follows:1.A research framework of anomaly structure change detection for dynamic networks is proposed,which mainly includes dynamic network feature extraction,dynamic network trend anomaly detection,dynamic network abrupt anomaly change detection and dynamic network anomaly link detection Four parts,which lays a foundation for subsequent research.2.The unremarkable trend anomaly change of network structure is a form of dynamic network anomaly change.The unremarkable trend anomaly change is difficult to be found by the method of identifying network abrupt anomaly change because of its small change amplitude.In view of this,this paper designs a dynamic network trend anomaly detection model by using the data trend detection method Cox-stuart test,combining the variable sliding window strategy to divide data.By introducing the variable sliding window,our approach can be more precise in identifying the time period when the network has a trend-changing anomaly.We use this method to detect the trend anomaly of dynamic networks on the real dataset,and the detected network anomaly change time is basically in agreement with the time of the real anomaly event,which verifies the validity of the method.3.The method of abrupt abnormal change detection in dynamic networks is more widely used in practice.Traditional methods of dynamic network abrupt change detection are mainly based on network structure characteristics,but it is difficult for us to extract accurate network structure characteristics in the face of large-scale and complex network data.To solve this problem,this paper proposes a method to detect the abrupt abnormal change of dynamic network by using network embedding.This method takes feature extraction as a embedding problem to automate the whole process and detects the abrupt abnormal change of dynamic network by analyzing the distribution of nodes in space after the network embedding process.The experimental results show that the network processed by network embedding is more effective than the original network in structure feature extraction.4.The intuitive understanding of abnormal link behavior is that two nodes that would otherwise be unlikely to have a link relationship have a link relationship,and these low-probability abnormal link behaviors are likely to be the cause of abnormal changes in the network.This paper designs a dynamic network abnormal link behavior detection method based on network embedding method.The degree of similarity between nodes after network embedding is used to describe the probability of link relationship between nodes.The link behavior with low probability is detected as abnormal link behavior.Our method is able to detect the proportions of the current time period abnormal link to all links and point out several link behaviors that are most likely to be outliers.The proposed method outperforms the traditional method in real datasets because it considers more and deeper implicit relationship between the nodes.