Research And Application Of Graph Double Attention Network Based On Maximum Pooling

Telecoms operators have huge amounts of call details,which reflect the social networks between people.The network contains not only the personal attribute information of the user,but also the network characteristic information of the individual.How to process and analyze such complex network data has become an important research content in the field of data mining.A graph is a data structure that models a set of entities and the relationships between them.In recent years,due to the strong expressive force of graph structure,the research of analyzing graph with machine learning method has been paid more and more attention.Network representation learning is a method of processing network or graph structured data.The idea of Graph Convolutional Network(GCN)based on network representation learning is to aggregate the local neighborhood and attribute information of nodes to generate node embedding.Graph convolutional neural network has better performance as a method of processing graph structure data,so GCN has become a widely studied and applied graph analysis method.Graph Attention Network(GAT)is an attention mechanism added on the basis of GCN,which enables neighborhood nodes to make effective information contribution to the central node,and ignores unnecessary information in the prediction to focus on the most relevant nodes around.This paper focuses on the model of graph convolutional neural network based on attention mechanism and the problem of customer churn in call data.The main research contents include the following aspects:(1)This paper proposed relevant analysis and data preprocessing on the call data,and it uses the call data to build a mobile social network.The change of network characteristics before and after customer churn is analyzed,and the attribute network model is constructed,which lays a foundation for the research of graph neural network on the classification of graph nodes.(2)This paper put forward a kind of graph neural network model based on double attention,which aimed at solving slow convergence speed of the training process and theaccuracy rate.In other words,the method of increasing vertical attention on the basis of the original graph attention network(GAT)can "enhance" attention,force the nodes to learn the feature contributions of different nodes to the intermediate nodes,and make the learned model more accurate and converge faster.(3)In view of the insufficient of existing graph neural network generalization ability,this paper is based on the particularity of graph structure proposed graph pooling operation.The operation can be embedded to any intention of graph neural network model,and the experimental results show that pooling operation could enhance the generalization ability of the model.From a certain extent,it solved some graph convolution smoothing problem of neural network,and improved the accuracy of the model.(4)The graph neural network model is applied to the real social network for the prediction of telecom customer churn.Compared with the traditional random forest,XGBoost method and GCN and other deep learning methods,it has significantly improved the evaluation indexes such as accuracy and recall rate.