| From social networks to the World Wide Web,as an important form of data,the network is widely used to describe the complex relationships between things in the real world.Effectively mining potentially important information from complex information networks has become a research hotspot in academia and industry in recent years.Network representation learning,also known as graph representation learning,graph embedding,the purpose is to represent the nodes in the network with low-dimensional,dense,real-valued vectors,and apply existing machine learning algorithms to perform a series of network analysis tasks,such as links Prediction,node classification,and node clustering,etc.At present,most of the representation learning of the network is only for the topology information of the network,but with the diversification of network information,the nodes in the real network usually contain a large amount of attribute information,and the learning of this information is very important for subsequent applications.This dissertation attempts to conduct related research from the perspective of improving the quality of attribute network representation.Aiming at the shortcomings of other methods that topological information and attribute information cannot be effectively combined,the main work and contributions of this thesis are reflected in the following three aspects:(1)Aiming at the problem that random walk cannot take into account the structural equivalence and homogeneity of attributed network representation learning,this dissertation proposes an attributed network representation learning method based on the biased random walk.Specifically,first,perform random walks between ordinary nodes and attribute nodes in the attribute bipartite graph.In the structural network,a bias parameter is introduced to balance the structural equivalence and homogeneity of different networks.Then,the node sequence generated by the random walk is compared with the sentences in the natural language processing corpus,and the node sequence is sent into the skip-gram model for attributed network representation learning.The experimental results on the real attribute networks show that the accuracy of the method in the node classification task has been improved.(2)Aiming at the insufficient enhancement of complementary structural information in the existing attributed network representation learning methods when learning attribute information,this dissertation proposes a structure-enhanced attribute network representation learning method to improve the quality of representation.More specifically,this method first extracts attribute information that enhances the global adjacency of the network through an autoencoder based on the normalized adjacency matrix and attribute matrix,then uses the contrastive Loss to capture local structure information,and finally introduces a joint loss function to make the structure information and attribute information represent in the same vector space.Extensive experiments of node classification and link prediction on three real attribute networks demonstrate a superior performance of the proposed method over state-of-the-art for network representation learning.(3)In view of the fact that the existing embedding algorithm based on graph convolution network not only ignores the data distribution of implicit representation but also loses the higher-order proximity between nodes in the graph,this dissertation proposes a novel graph embedding framework based on adversarial and random walk regularization,which jointly preserves structural and attribute information.This method adopts an adversarial training scheme to enforce the latent codes to match a prior distribution,and by employing the loss function of skip-gram model,nodes in a random walk sequence are closer in the latent space.Experimental results on link prediction and graph clustering task on the real attributed network datasets demonstrate that this method achieves better performance than state-of-the-art graph embedding algorithms. |
