Research On Pooling Algorithm Based On Graph Neural Networ

Deep learning algorithms based on graph neural networks are currently attracting widespread attention in the field of graph classification.The pooling operation is a relatively common operation in graph neural networks,and it is often inserted after convolution,which can achieve the purpose of reducing the dimensionality of the vectors output by the convolution layer and reducing the network parameters,as well as avoiding the effects of overfitting,etc.However,the currently proposed graph pooling algorithm suffers from insufficient information consideration,inadequate modeling,poor classification accuracy and other problems.This paper studies graph neural network-based pooling algorithms to fully consider the node feature information and structural information in the input graph to mine the potential information of the graph and achieve a better graph representation.Based on this,this paper does the following three aspects:In the first part,the adaptive graph pooling algorithm SFMPool is proposed to address the problem of incomplete consideration of the information in the graph in the graph classification task.the algorithm fully considers the node feature information and the structure information of the graph in extracting the information in the graph,and obtains the potential information existing in the graph based on two levels,feature level and structure level,respectively,to complete the graph classification task.In the process of this model optimisation,the algorithm proposes to increase the dependencies between the input graph and the pooled graph through a mutual information mechanism to enhance the representation of nodes in the graph.The algorithm is able to mine the potential information contained in the graph and thus obtain a more realistic representation of the graph.Experimental results with real data sets show that this algorithm has some performance improvement over existing pooling methods.In the second part,a graph pooling algorithm based on structure enhancement is proposed in order to further optimise the pooling algorithm and to deeply mine the structural information contained in the graph.This process extracts the structural information of the graph by the importance of the nodes,after which a structure learning method is designed to enhance the connectivity of the graph and greatly reduce the rate of information loss.Finally,an attention mechanism is added to improve selfexpressiveness,which both strengthens the structure and preserves the node information in the graph,substantially improving the accuracy of the graph classification task.In the third part,to address the problem of single,one-sided consideration of the model in the pooling process,a dual-driven graph pooling algorithm is proposed,which solves the problems encountered by most of the methods in the pooling process by integrating the advantages of cluster pooling class methods and node selection class pooling methods,while ensuring accuracy.Two pooling graphs are first obtained by parallelizing two pooling channels which focus on different aspects of the graph,and then a convolution method is proposed to process the pooled graphs to obtain the final pooling results.Experiments show that this method significantly improves performance in all aspects compared to other pooling methods.