Virtual Cyberspace Information Mining Baced On GNN

Machine learning has achieved remarkable performance on Euclidean spatial data such as images,text,and speech,while there are relatively few studies on non-European spatial data.However,non-European spatial data represented by graph network data exist in all aspects of life,such as social relationship network data,traffic data and molecular structure in the field of biochemistry.Information mining tasks based on graph network data mainly include graph classification,link prediction and node classification.However,since the graph network data does not satisfy translation invariance,traditional machine learning models are not suitable for such tasks.Therefore,the study of machine learning algorithms suitable for graph network data has become a research hotspot in recent years.In the early era of manual feature extraction,researchers obtained feature information on graph data according to rules based on statistical analysis methods.Up to now,there are many machine learning algorithms and deep learning algorithms based on graph neural network.However,as the number of network layers increases,there are still some urgent problems to be solved,such as the over-smoothing problem that makes the nodes representation inseparable.This dissertation proposes two different solutions based on solving the problem of over-smoothing during training of deep graph neural networks.One of them is based on the improvement of the existing research,and the BernNet has been expanded.The other proposes BoostResGCN,a new deep graph convolutional neural network framework.Two kinds of extension studies are carried out on the original BernNet.First,considering the non-differentiation of the ReLU function used by the original BernNet for coefficient non-negative constraint at zero points,we use the square function of the ReLU function to replace the ReLU function.Experiments show that the squared model achieves the best performance on multiple datasets.In addition,BernNet’s non-negative constraints on model coefficients will have problems such as insufficient model performance.Therefore,we remove the non-negative constraint on the coefficients.Experiments show that even without coefficient constraints,the performance of the model will not be much different from the performance of the model with non-negative coefficient constraints,and even better on some datasets.For the deep graph convolutional neural network framework BoostResGCN,we use the graph convolutional neural network GCN as the residual block and apply the boosting theory to build the network framework.The training model adopts a layer-bylayer training strategy,so the classifier of the final model is equivalent to integrating multiple weak classifiers.We also explore different ways of connecting the residuals.Experiments show that BoostResGCN achieves better classification results than the depth graph convolutional neural network without the Boosting theoretical framework.The residual connection using the addition is relatively stable and takes less time.BoostResGCN has achieved certain results in combating the problem of oversmoothing.