Research On Graph Classification Based On Graph Neural Networks

With the improvement of the degree of informatization in various industries,the data presents the characteristics of structure,diversification,and complexity.As a general data structure,graph can model complex data generated in the real world.After modeling,many application scenarios are transformed into mining potentially useful information from graph data,namely graph mining.Among those graph technologies,graph classification,an important branch of graph mining,has received extensive attention.At the same time,CNNs(Convolutional Neural Networks)have played a huge role in computer vision,natural language processing and other fields.Some researchers have extended the neural network to graph data and proposed GNNs(Graph Neural Networks),hoping to use the powerful feature modeling capabilities of neural networks to solve graph classification problems.Although research on GNNs and their applications in graph classification have been carried out,there are still many problems to be solved.For example,most hierarchical pooling methods empirically use simple permutation invariance functions such as Mean and Max to obtain the representation of the graph in the hierarchy.Such methods default that all nodes are equally important,ignoreing the differences in nodes' significant,resulting in a large loss of key node information and non-key node information interfering with the graph classification.And most existing studies focus on improving the feature modeling capabilities of graph neural networks,designing more complex graph convolution operations,but ignoring the computational overhead caused by iterative graph convolutions,resulting in slower network training speed.Based on the above problems,this paper studies graph pooling methods and simplified graph neural networks,and proposes corresponding solutions.The specific research content is as follows:(1)We propose a weighted graph pooling method WGP,which learns the weights of nodes based on their topology information and node label information to distinguish the differences in nodes' significant.Then WGP calculates the weighted sum of all node representations as to the representation of the graph according to the node weights.Compared to permutation invariance functions,WGP can utilize the importance of nodes and effectively extract the key features of the graph.WGP can also be integrated into a variety of GNN frameworks to optimize in an end-to-end fashion.The experimental results on the benchmark datasets show that WGP can further improve the accuracy of graph classification compared to state-of-theart graph classification methods.(2)We propose a succinct graph neural network SGNN for graph classification and prove that it is a linear approximation of GCN.SGNN pre-calculates multi-scale local smoothing of nodes by removing non-linear transformations in the process of graph convolution,and then designs a set-oriented neural network model to perform linear graph convolution and graph pooling of multi-scale node features.Compared to GCN,SGNN avoids the computational overhead caused by the iterative process of nonlinear graph convolution and reduces the amount of network computation.Experimental results on the benchmark datasets show that SGNN can effectively accelerate the training of networks while maintaining classification accuracy comparable to state-of-the-art graph classification methods.