Research And Application Of Explainability Algorithm In Graph Neural Networks

Although graph neural network is effective and widely used in many fields,its complex structure makes it difficult to explain the algorithm but the explanation is critical to the verification.Although there have been explanations for graph neural networks,the existing explain models fail to attribute the decision-making results fairly to the factors,and only evaluate the contribution of nodes,which is difficult to evaluate the subgraph-level contribution and ignores the interaction between nodes.Thirdly,although a few can explain at multiple granularities,the complexity is unacceptable.In order to solve the above challenges,we first proposed a subgraph-level graph neural network SUPExplainer based on reinforcement learning.Without prior knowledge,this model selects important subgraphs effectively to represent the key structural information of the graph by strengthening the learning mechanism,and to maximize the cumulative reward by clipping near-end strategy optimization(ClipPPO)in the explanatory environment.In order to evaluate the contribution of subgraphs fairly and effectively,Shapley value is regarded as a reward function when designing the reinforcement learning framework,and Monte Carlo sampling method is used to improve the sampling efficiency and reduce the calculation amount.In this thesis,ablation experiment,quantitative experiment and statistical test are used to test the effect of the model on graph classification and node classification data sets,and convergence analysis and efficiency analysis are also carried out.The results show that this method can not only display the important graph structure in the interpretation process visually,but also outperform some baseline methods in reasonable time complexity.And it is significant and effective at the significance level of 0.05.It can provide a reliable explanation for GNN model,and it is also referred to the application of reinforcement learning in the field of graph interpretation.Finally,to verify the explanatory effect of GNN in the text classification task,the thesis compares the text classification effects among Text GCN,bert GCN and textlevel GCN through SUPExplainer.The experimental results show that the interpretation model can intuitively understand the text classification effect,facilitate a better application of deep learning methods.