Semi-supervised Node Classification Base On Hypergraph Neural Networks

With the rapid development of various technologies,data has exploded in all aspects of people’s life,transportation and entertainment.The classification management of data has become an urgent need in people’s daily life.Semi-supervised classification is one of the most popular research problems in machine learning.It requires only a small amount of labeled data and a large amount of unlabeled data to achieve good classification performance,effectively reducing the manual labeling cost of supervised learning and the inaccuracy of unsupervised learning model.In recent years,semi-supervised node classification based on hypergraph has become an important branch of semi-supervised learning,which utilizes the advantage of hypergraph to model high-order correlations among data,and realizes the transmission of label information via hypergraph nodes.This paper aims to implement the semi-supervised node classification task using hypergraph convolutional neural networks.A hypergraph pooling convolutional neural network and a neural network combined with graph convolution and hypergraph convolution are designed with the objectives of improving classification accuracy and reducing model complexity.The main research results are as follows:(1)To address the problem that the existing hypergraph convolutional neural network is shallow and lacks the processing of redundant information,which makes it difficult to extract features due to the small receptive field of the learned nodes,a hypergraph pooling convolutional neural network is designed to implement node classification.Firstly,the hypergraph structure is transformed into a graph,so that the graph pooling operation can be extended to deal with hypergraph data.Secondly,a hypergraph U-Net network consisting of multilayer encoding unit blocks and decoding unit blocks is designed to extract discriminative features.The encoding unit block consists of a hypergraph pooling layer and a linear encoder to capture depth features by expanding the receptive field of nodes through down-sampling.The decoding unit block consists of a hypergraph Unpooling layer and a linear encoder,which is used to recover the topology connectivity of the network and facilitate the propagation of node information.Finally,the multi-layer perceptron is used to obtain the final prediction of nodes.Semi-supervised node classification experiments are conducted on the citation network dataset,and the experimental results show that the proposed model can improve classification accuracy.(2)To address the problem of noise interference introduced by existing hypergraph neural networks using multilayer hypergraph convolution and the high computational complexity of using multichannel models,a neural network learning framework combined with graph convolution and hypergraph convolution is designed.Firstly,the single-layer hypergraph convolution is used to reduce the noise interference caused by multi-layer convolution in node neighborhood expansion and extract the high-order correlation of the data.Secondly,the hypergraph is transformed into a graph to construct the spatial structure suitable for graph convolution operator.Finally,graph convolution is used to further fuse the information of higher-order correlation and graph topology information to enhance the representation of node features.The performance of the model is evaluated on three citation network datasets,and the experimental results demonstrate that the proposed model not only effectively improves the classification accuracy but also reduces the computational complexity.