Multi-granularity Complex Network Representation Learning Based On Random Walk

Network representation methods are generally divided into two types.Traditional topology-based network representation method usually uses adjacency matrix,which may contain noise or redundant information.Embedded-based network representation method is designed to learn dense and continuous representations of nodes in low-dimensional space,which can reduce noise or redundant information and retain inherent structural information.Since each node is represented by a vector containing the information of its interest,many problems in network analysis can be solved by calculating mapping functions,or distance metrics,thus avoiding highly complex operations.Most of the existing embedding methods can only learn about one of the properties of the network.Taking homogeneity as an example,methods that learn this property usually perform well in link prediction and network reconstruction tasks.There are also some methods used to capture the isomorphism of the nodes,which generally perform well in the task of node importance classification.This thesis proposes multi-granularity complex network representation learning based on random walk,which can simultaneously capture the isomorphism and homogeneity in the network.The main innovations of this thesis are as follows:A multi-granularity network representation learning method based on game theory is proposed.First,divide the network into multiple granularities according to the global structure characteristics,calculate the similarity matrix on each granular layer,and construct the mapping relationship between the granular layers.Then combined with the idea of propagation dynamics,the nodes in the network are used as game individuals to realize dynamic random walk by constructing the income matrix.Finally,the natural language processing model Skip-Gram is used to train the node sequences,adjust the parameters by maximizing the probability of node co-occurrence to obtain low-dimensional vector representation with semantic information.A fuzzy hierarchical network embedding method based on isomorphism and homogeneity is proposed.First,in order to better capture the node importance semantics in the network,the fuzzy k-core decomposition method is proposed.Then,based on the epidemic model,the random walk process is analogous to the process of information dissemination,generating node sequences with biased random walk on multi-granularity graph.Finally,according to the statistical information of the global node co-occurrence,the Glove model is used to learn the vector representations of the nodes.Experimental results show that the models proposed in this thesis have performed well in many experiments.In many experiments such as classification,clustering,link prediction,and visualization,the models in this thesis are superior to the comparative method in many indicators.Finally,this thesis takes a social network as an example to analyze the model's ability to process the actual network.