Research On Network Representation Learning Based On Multi-Granular Structure

With the advent of the era of big data,social networks,etc.are rapidly developing and generating massive and complex network data and implying a lot of important and valuable information.In the face of massive data,the traditional adjacency matrix and other vector representations have problems of sparse vectors and high computational complexity,while network representation learning converts massive data into low-dimensional dense vector representations and uses them as input to commonly used machine learning algorithms.Carrying out network analysis tasks makes it possible to perform network analysis quickly and efficiently,so it has important research significance.The core idea of network representation learning is to find a mapping function based on preserving the topological structure characteristics of the network,and convert the nodes in the network into low-dimensional dense representation learning vectors,which can be used in subsequent network analysis tasks.The existing network representation methods based on a single granularity structure mainly include a representation learning method based on matrix decomposition,a representation learning method based on random walk,and a representation learning method based on deep neural networks.However,a series of studies on complex networks have confirmed that many networks in the real world exhibit a multi-granular structure,and the use of random walk to mine the structural characteristics of the network exists to capture only low-order structures while ignoring the higher-order structural characteristics of the network the shortcomings of this study,on the basis of retaining the local structure of the network,this study introduces multi-granular structure features to improve the accuracy of network analysis tasks.In summary,the main research work of this thesis includes:1.This article first gives a brief introduction to the background knowledge of online representation learning,and fully investigates relevant domestic and foreign networksThe network represents the research status of learning problems and basic method theory.In addition,we focus on analyzing the main difficulties and challenges faced by the network representation learning problem,and propose corresponding research methods for these challenges.Therefore,to better understand the complex network and explore the potential laws of the complex network for better application,this study is dedicated to the mining of the multigranular structure of the complex network and proposes a multi-granular community structure network representation learning method and the network representation method of granular high-order structure.2.With the deepening of experts and scholars in the field of complex network research,more and more studies have proved that real complex networks often exhibit the characteristics of multi-granularity structure.To solve this problem,this research is based on the multigranularity community structure and proposes the network representation learning method of granular community structure HCNE.We construct a multi-granularity network from fine to coarse through network granulation to obtain the multi-granularity community structure of the original network.To retain the characteristics of the multi-granularity community structure,we use the vector representation of the coarse-grained layer as the initial value of the vector representation of the previous fine-grained layer to implement the inheritance of feature information between the coarse-grained and fine-grained layers;to retain the local structure,we use random walk the model stores local structural feature information at multiple granularities.Finally,experiments were performed on multiple data sets.The results show that the representation learning method proposed in this study has certain applicability in network analysis tasks.3.Due to the use of random walk to mine the structural feature information of the network,but constrained by the length and scope of the walk,only the low-order structural features are captured and the high-order structural feature information of the network is ignored.To solve this problem,a network representation learning method based on multi-granularity high-order structure is proposed.First,the complex original network is simplified by merging nodes and edges,and a series of multi-granular networks with fine to coarse granularities are obtained.Then,by using the vector representation of the coarse-grained layer as the initial value of the vector representation of the previous layer,the multi-granular structure of the network is retained during the process of training and learning to form the network embedding.Considering the network representation model based on random walk,there is a global structure that cannot effectively capture the network due to the constraints of the length and scope of the walk,so we introduce a representation learning method that can retain the higher-order structure of the network more effectively Disadvantages.Finally,we perform nonlinear training on the two-part model,and fuse the vector representations obtained from the two parts as the final vector representation.Finally,experiments are performed on multiple data sets,and the results show that the performance of this study on network analysis tasks has been improved to a certain extent.