Exploring The Modeling Of Dynamic Community Evolution For The Node Transition Behaviors

It is the core connotation of the framework of intelligence analysis to use multilayer complex network for sensing,understanding and predicting.Among them,event detection,anomaly detection and relationship prediction are the main contents of current intelligence analysis.Based on the models of community evolution in dynamic complex networks,the dynamic community structures and the corresponding community evolution patterns can be digged.Furthermore,the anomaly detection of community evolution can be applied to event detection in real social networks.Dynamic community evolution is usually driven by the node transition behaviors among communities.Thererfore,the emphasis of modeling dynamic community evolution is how to quantitatively characterize nodes’ transition behaviors,and the difficulty is how to accurately depict the inconsistencies of nodes’ transition behaviors.In this thesis,aiming at the inconsistencies of the node transition behaviors,we focus on studying the methods for modeling the dynamic community evolution.The main contents and innovative results are as follows:First,focusing on the quantification of the nodes’ transition behaviors,the community evolution matrices used to quantify the nodes’ transition behaviors is introduced based on the classical evolutionary clustering framework,and the constraint of the historical local information of nodes is integrated in the form of graph regularization synchronously.And then a scalable dynamic community evolution framework based on graph regular nonnegative matrix factorization(NMF)including the corresponding optimization algorithm is proposed.It can effectively improve the accuracy of dynamic community structure detection.Second,aiming at overcoming the difficulties from the inconsistency of nodes’ transition trend,we design a manner using an easily extensible similarity index for quantifying the first-order varying information across snapshot of nodes,and introduce it to our model to eliminate the effects on community evolution from the inconsistencies of nodes’ transition behaviors.And then a robust framework including the corresponding optimization algorithm for modeling dynamic community evolution is proposed by fusing the firstorder varying information of nodes.Third,aiming at solving the problem that the number of communities to which nodes turn is uncertain,we add a priori distribution for the community evolution matrix based on the classical bayesian nonnegative matrix factorization(BNMF).And then a dynamic community evolution model including the corresponding optimization algorithm are proposed from the perspective of model selection,which automatically determine the number of dynamic communities of each network snapshot.At last,addressing the problem that the nodes’ transition patterns are unstable,we model the dynamic community evolution from local and global perspectives in a two-step strategy,and propose a new model including the corresponding optimization algorithm based on orthogonal nonnegative matrix factorization(ONMF).And we compute the intensity of community evolution with the obtained local and global community evolution patterns,which can be used to detect the anomalies of dynamic community evolution.This thesis focuses on the inconsistency of node transfer behavior,mainly studies the modeling method of dynamic community evolution,mines the dynamic community structures and its evolution patterns,and applies it to the event detection in the real social networks.