Research On Technology Of User Behavior Modeling Based On Graph Mining

Social networks and Internet communications have gradually become an important part of people's lives,however,huge risks are hidden behind great conveniences.Malicious users have brought great dangers,to deal with this problem,user behavior modeling analysis based on graph has gradually become a hotspot in academic at home and abroad.This thesis analyze abnormal behavior detection in static graph and dynamic graph,and put forward corresponding solutions,as follow:(a)In anomaly detection of static graph,an unsupervised anomaly detection algorithm KD-Forest is proposed.Extracting feature based on graph structure information and using KD tree to construct classification tree,and Bagging method is used to select samples to improve randomness.Experiment on more than one hundred thousand social network dataset prove that our algorithm has good scalability.Compared with the existing related detection algorithms,the proposed method has higher time efficiency,space efficiency,better accuracy and ROC performance.(b)In anomaly detection of dynamic graph,we propose an algorithm based on LSTM to detect time series anomaly.Graph similarity measurement methods based on graph distance are proposed,including structure distance and edit distance.Training classification model with these features and experiment on dataset of over one million network data stream,this method has higher accuracy and ROC than the existing algorithms.These two kinds of algorithms proposed in this thesis are suitable for the anomaly detection of social network and communication network.Corresponding solutions are proposed according to the type of graph,and obtain better detection accuracy.Finally,the existing shortcomings and future research areas are discussed.