Structure Identification Of Complex Networks Based On Group Feature Selection Nonlinear Granger Causality

In the current era of big data,the research on complex networks has brought enormous economic and social benefits to human society.Among them,how to identify complex network topology based on observations of nodes has become a focus in multidisciplinary research.Indepth study of network structure identification is helpful to understand the function and behavior of complex networked systems and grasp the emerging characteristics of network structure.Based on the classical Granger causality network model and with the consideration of the nonlinear and sparse structural characteristics of complex networked systems,this thesis puts forward two nonlinear Granger causality methods based on group sparse feature selection.The main work of this thesis is summarized as follows:1.The nonlinear conditional Granger causality algorithm based on block orthogonal matching tracking(BOMP-NCGC)is proposed.Firstly,based on the classical model of complex networked system,the nonlinear conditional Granger causality model is constructed.Then the block orthogonal matching tracking algorithm is used to solve the sparse selection problem of group feature relationships among nodes.With the comparison of the classical conditional Granger causality method(CGC),the simulations on different network types verify the effectiveness of the proposed method.2.Group sparse least squares nonlinear condition Granger causality(GSLS-NCGC)is proposed.Firstly,the multiple nonlinear conditional Granger causality model is constructed,and the kernel function is used to fit the nonlinear relationship between nodes.Then,the group sparse penalty least squares method is used to select the corresponding group features and construct the final topology.The sparsity of inter-group relationship is considered,and the nonlinear overfitting problem within the groups is solved.Through the simulation experiments on different network types,network sizes and model parameters,The effectiveness and robustness of the proposed method are demonstrated.