Research On Graph Neural Network Learning And Application Based On Graph Structure Optimization

Graph neural networks are an important method for analyzing graph data,aiming to utilize the topological structure of the graph to map high-dimensional,dense feature of nodes in the graph into low-dimensional vector representations.The excellent ability of graph neural networks to depict complex association relationships and interactive behaviors between entities in real systems has led to breakthroughs in various tasks such as node classification,link prediction,and recommendation systems and has become a hot research direction in the field of data mining.At present,most graph neural networks are used to directly represent graphs in learning reality,without considering the possibility of graph interference.However,missing,meaningless,or even false association relationships are common in graphs,indicating the uncertainty of the graph structure.Two network nodes with relatively strong association relationships may not have connections in the graph,while two network nodes with relatively weak association relationships may have connections.In addition,real systems may experience changes in their association relationships due to adversarial attacks and other factors.The existing graph neural networks have a significant decrease in predictive ability when faced with interference in the graph structure,exposing the problem of weak robustness.Therefore,the study of representation learning modeling for graph data with node relationship uncertainty is an urgent scientific problem that needs to be solved,with profound theoretical value and broad practical application prospects.This thesis conducts in-depth research on graph structure optimization in graph neural networks.The main research content and innovation points are as follows:(1)Aiming at the uncertainty of node relationships in a graph,this thesis proposes a Bayesian Graph Attention Network model.The model first uses a graph attention network to calculate the weight of the target node’s neighborhood,learn to generate more detailed node information,and generate multiple observation graphs based on multi-level node information as observation information for optimizing the graph structure.Then,the model uses Bayesian inference to re-estimate the topological structure of the graph,providing a more realistic graph structure for graph representation learning and achieving mutually beneficial optimization of graph representation learning and graph structure estimation in an iterative framework.The experimental results show that the model has good classification performance.(2)Aiming at the problem of the weak robustness of graph neural networks caused by graph antagonism attacks,this thesis proposes an Evolutionary Graph Neural Network model.The model first learns multiple graph structures as mutation factors and mutates a set of graph neural network model parameters to adapt to the environment(that is,improve the classification performance of unlabeled nodes).Then,an evaluation strategy is proposed to measure the quality of the generated samples,preserve model parameters with good performance(as descendants),and preserve the descendants that adapt to the environment for further evolution.The experimental results show that the model has good classification performance and strong robustness.(3)Aiming at the problem of poor homogeneity in the topological structure of Ethereum trading networks,which is difficult to accurately characterize using traditional graph neural networks,this thesis proposes an Adaptive Multi-channel Bayesian Graph Attention Network model.Firstly,this thesis explores the characteristics of the network and finds that the homogeneity index of the network is low,which affects the representation learning ability of the graph neural network.Then,Bayesian inference is used to estimate the topological structure of a graph that conforms to ground truth.A multi-view strategy is used to extract respective and common embedding information from the estimated graph,the original graph,and their combinations.Attention mechanisms are used to adaptively learn the importance weights of the three types of node embedding.The experimental results show that the model has good classification performance in the Ethereum fraud account detection task.In this thesis,Bayesian inference and an evolutionary algorithm are used to optimize the graph structure and improve the representation ability of a graph neural network.In addition,this thesis utilizes an adaptive multi-channel Bayesian graph structure optimization method to improve the classification performance of graph neural networks in the domain of fraudulent account detection in Ethereum.Finally,this thesis makes a reasonable case for future research work.