A Study On Graph Pooling Method For Graph Classification

Graph neural networks can efficiently process graph data and have a wide range of applications in the fields of social media,bioengineering,and transportation.The task of graph classification is an important application of graph neural networks,which requires designing a pooling layer to obtain the graph-level representation of graph data.Therefore,this thesis designs a series of efficient graph pooling methods on the basis of existing neural network methods,which show superior performance in graph classification tasks.The specific works are as follows:1.A spectral graph pooling method based on graph wavelet transforms is designed.This method uses graph wavelet transform to map the graph embedding obtained by the graph convolution layer to the spectral domain and selects the sampling node through the importance score learned by graph spectral components.In particular,two kinds of graph wavelet transform that allows fast implementation is used in our pooling framework.Technically,the proposed framework is generic in the sense that different spectral graph transforms can be used to design the pooling operation.Experiment results show that the proposed spectral graph pooling method has good potential in producing better graph-level learning representations.2.A multi-head densely connected graph pooling model combined with structure learning is constructed.The model learns the node importance scores through a multi-head densely connected network,sets different convolution weights for each network,and then implements different filtering operations.Weight sharing within the network so that no additional complexity is added when increasing the number of network layers to extract high-order features.At the same time,using dense connections,the outputs of all previous layers are passed to the next layer,and the initial features are added to each layer,which enhances feature transmission and efficient use of features.Then the graph structure learning module is introduced to learn the correlation degree for each pair of nodes in the pooling graph.If the correlation degree between the two nodes is high,they are connected to form a new edge,so as to ensure the connectivity of the whole graph.Experimental results show that the proposed method exhibits superior performance in graph classification tasks.3.A multi-head attention graph pooling model combined with information retention is built.This model adopts the multi-head attention mechanism,calculates the correlation between the central node and the neighbor nodes multiple times,and learns the node importance score more comprehensively.Meanwhile,the information retention module is built to retain effective information in the discarded nodes.Finally,the isolated nodes in the pooling graph are connected with the neighbor nodes to ensure the connectivity of the graph structure.The experimental results show that the proposed model can produce better results in graph classification tasks.