Research On Text Classification Algorithm Based On High-order And Low-order Graph Convolutional Network

With the rapid development of mobile Internet technology,the number of network users has increased dramatically,and tens of thousands of connected text network data have been generated.Using deep learning technology to dig out the hidden internal rules of the network from the complex network structure,which has important theoretical significance and application prospects for analyzing and studying the network behavior and the hidden connections between network nodes.We use graph convolutional network(GCN)to deeply research and analyze text classification.We propose several high-order and low-order graph convolutional network models to improve the accuracy of text classification.The main work of this paper is as follows:In terms of capturing text node information,aiming at the limitation of GCN with shallow mechanism and over-smoothing problem of multi-layer(>2 layers)GCN,we design a highorder and low-order graph convolution based on weight sharing mechanism to simultaneously capture the high-order and low-order neighborhoods information of text nodes.In order to fuse the high-order and low-order neighborhood features extracted by different order graph convolutions,we propose four information fusion schemes,including maximum information fusion pooling,mean information fusion pooling,sum information fusion pooling,and inverse minimum information fusion pooling.According to the four proposed information fusion schemes,we propose four corresponding text classification algorithms based on high-order and low-order graph convolutional networks.In addition,we analyze the computational complexity and the amount of parameters of these models.Experimental results show that the proposed high-order and low-order graph convolutional networks can effectively capture the first-order and high-order neighborhood information of text nodes and learn the global graph structure.In order to avoid the overfitting of high-order and low-order graph convolutional networks,we further propose a text classification algorithm based on graph convolutional network without hidden layers.The core of the text classification model lies in the proposed multi-step graph convolutional layer.Multi-step graph convolution includes three stages: multi-power adjacency matrix,multi-step feature propagation and multi-step linear transformation.Compared with high-order and low-order graph convolutional networks,the network can learn more and richer text node features and improve the performance of text classification.