| A network is an efficient data structure for describing entities and inter-entity community relationships.In real world,there are various information networks with large-scale and complex relationships,such as academic citation networks,social relationship networks,bioinformatics networks,etc.In-depth mining and analysis of these information networks have important academic research significance and practical application value.Network representation learning is a key technique for analyzing complex information networks,which obtains low-dimensional vector representations of nodes by maximizing the feature learning of the original network and can be applied to downstream network analysis tasks such as node classification,clustering,and visualization.Most of the existing network representation learning algorithms focus only on retaining the structural information of the network,but the nodes in real networks often have a large amount of attribute information,which helps to improve the quality of node representation.In addition,the existing methods do not take into account the community information within the network and ignore the higher-order implicit relationships between nodes and the potential clustering structure of the network.To solve the above problems,this paper tries to fuse network topology,node attributes,and community information to obtain a representation of the nodes.The main research content is as follows:(1)To address the problem that existing algorithms pay less attention to the rich attribute information of nodes and the local community properties of nodes,in this paper,we propose a multi-view network representation learning algorithm based on local community enhancement(MV-LCE).The algorithm preserves the topology and node attributes of the network from both structure and attributes views simultaneously.To enhance the interaction between the views,an attention-based multi-view fusion strategy is designed to encode consistent and complementary information from structure and attributes to achieve the fusion of information between the two views;and to reconstruct the average attributes of the local community in which the nodes are located using a decoder to make the node representations in the local community more similar.The experimental results of node classification and visualization on five real datasets show that the node representation learned by the MV-LCE algorithm outperforms the existing classical algorithms.(2)To address the problem that existing algorithms ignore the global community structure of the network,this paper proposes a network representation learning method based on dual-channel homogeneity matrix(NDHM).The algorithm firstly obtains the network homogeneity matrix in structure space and attribute space respectively;secondly,integrates the homogeneity matrix of two channels to obtain the overall network homogeneity matrix,and selects the nodes with higher homogeneity for each node to construct a new network with higher homogeneity;finally,aggregates the features of neighboring nodes based on the homogeneity weighting to obtain the node representation.The results of node classification,clustering,and visualization experiments on five real datasets show that the NDHM algorithm has good node representation learning performance.In order to preserve both the node local community structure and the network global community structure,two algorithms are fused to form LCE-NDHM,and experiments are conducted in five datasets to compare and analyze the three algorithms. |
