Research On Network Analysis Methods Based On Model Fusion

Network science models the real-world complex systems into the form of complex networks.A variety of network analysis methods have been proposed for network information exploration.Existing network analysis methods can be divided into two major categories: deep learning-based methods and probabilistic graphical model-based methods.Under each category,there are sub-categories based on different technologies.Because of different technical characteristics,different types of models have their specific advantages and disadvantages.Compared with using single network analysis model,model fusion has two main advantages: first,breaking the inherent disadvantages of a single model;second,mutual advantages complementation of various models.Therefore,this paper focuses on the topic of "network analysis with model fusion",and carries out the following two works from the perspectives of improving model shortcomings and integrating model advantages.As a rapidly-developed deep learning model,Graph Convolutional Neural Network(GCN)has been widely studied and applied in the academic and industrial fields.However,the GCN propagation and aggregation mechanism is restricted to homophily assumption(where similar nodes tend to be connected),making it hard to be used for analyzing heterogeneous network(where dissimilar nodes tend to be connected),such as protein structure networks,dating networks,etc,and therefore resulting in performance degradation.To solve the GCN homophily problem,this paper proposes an integrated network analysis method that combines directed probabilistic graphical model and graph convolutional network.To improve the propagation and aggregation mechanism,this method introduces the block model(a directed probabilistic graphical model)to learn the homophily or heterophily of the connecting regularities of various classes in the network.Besides,the method innovatively introduces a learnable block similarity matrix to capture the interrelationships of connecting regularities between nodes of various classes.Then,this method guides the GCN message propagation and aggregation based on the block similarity matrix,so as to ensure that the propagation and aggregation mechanism can be adjusted adaptively whether with homogeneous or heterogeneous networks.The new propagation and aggregation mechanism is not restricted by topological limitation,and can adaptively aggregate the information of homogeneous and heterogeneous neighbors.This paper compares this method with six other representative methods,and the results demonstrate the superiority of this method in both homogeneous and heterogeneous networks.In addition,from the perspective of integrating model advantages,this paper proposes a network analysis method combining directed and undirected probabilistic graphical models with attention mechanism.These two models have distinct advantages in modeling network characteristics.This paper transforms the two types of models into a unified form of factor graph,and designs a new attention layer to establish an end-to-end information exchange between the two models.In this way,two models perform their duties in their fields of expertise,while attention mechanism regulates the information fusion in the meantime.This paper compares this method with nine other algorithms in eight datasets,validating the superiority of this method.Furthermore,this paper verifies the rationality of the motivation of this method through case analysis.