Research On Dual-channel Graph Convolutional Network Based On Mixed Features

In real life many scenarios can be modeled with graph data,such as social networks,etc.Due to the ubiquity of graph data,the study of graph data has the great significance.Analytical methods of the traditional graph data are mostly mined based on statistics or artificial extraction.In most cases,they perform poorly and the computational complexity is too high.In the coming years,the rapid development of convolutional neural networks has attracted widespread attention because of its powerful modeling capabilities.Therefore,the use of convolutional neural networks to process graph data has become a hotspot in graph research,and the concept of graph neural networks has emerged.Graph convolutional networks have been successful in the fields of recommendation systems.However,for the complex information contained in the network,effective methods are needed to extract useful information from it.The existing single-feature graph neural network cannot completely characterize the relevant characteristics in the network.In network representation learning tasks,there are many studies on the joint modeling of text features and structure features through shallow neural networks.The results prove that hybrid feature modeling can perform better in downstream machine learning tasks.In graph neural networks,there is very little work on joint modeling of structure and text.Drawing on the network representation learning algorithm of joint text modeling,the performance of structure and text modeling in the graph neural network can be improved.In addition,due to the emergence of deep convolutional neural networks such as residual networks,the effect of tasks such as image classification has been greatly improved.The graph convolutional neural network increases the number of graph convolutional layers to occur the phenomenon of over-smoothing.And its feature expression ability does not increase but decreases.To avoid this phenomenon,thesis uses residual connections to construct a deep dual-channel graph convolutional network.Therefore,the research work of thesis is as follows:(1)A dual-channel graph convolutional network model based on mixed features is proposed,and the graph convolutional network is used to solve the problem of joint modeling of structure and text features.And draw on attention mechanism and gate unit to design aggregation function,selectively aggregate the text features of nodes,and enhance the feature expression ability of nodes.In order to make the features complement each other,a vector fusion mechanism is used to generate mixed features for the node classification task.Through comparative experiments and ablation experiments,it is proved that the graph convolutional network based on mixed feature training has better node classification performance than the single-feature graph convolutional network.(2)By migrating the residual connection to a two-channel graph convolutional network based on mixed features,the deep structure of the graph convolutional network is constructed,so that the nodes can transfer feature information with weakly related nodes far away to avoid the over-smoothing problem.Through comparative experiments,it is proved that the use of residual connections to build a deep graph convolutional network is more effective in node classification tasks,and the depth of the graph convolutional network can reach 16 layers.The research in thesis confirms that the graph neural network can learn useful features from the text of the nodes,improve the learning performance of the model,and can also provide theoretical basis and technical support for subsequent graph neural network applications.