Research On Convolutional Neural Network For Graphs

With the improvement of computer calculation ability and the development of deep learning theory,the convolutional neural network(CNN)has become a shining light in the fields of computer vision and natural language processing.However,the traditional CNN is limited to deal with Euclidean structural data,so it is difficult to play a role in graph data directly.At the same time,graph,as a general data structure,can flexibly model the data and the complex relationship between the data.Therefore,experts hope to extend CNN to graph data to solve practical problems such as node classification and graph classification.Although some related work has been done to extend the CNN to graph data,there are still many problems.In the process of constructing local neighborhood,only the topological structure information of the graph is considered when selecting the central node,while the feature information of the node is ignored.Another problem is the loss of node information caused by the normalization of local neighborhood.In order to solve the above problems:(1)We proposed a method to assembly local neighborhood called attention neighborhood assembly(ATNA)based on attention mechanism.Firstly,according to the graph topology information and node characteristics,ATNA learns a scalar for each node,which size reflects the importance of the corresponding node in graph.Then ATNA selects a fixed number of nodes from the nodes sorted by the value of scalar.Finally,ATNA uses breadth-first search to build local neighborhoods centered on those nodes.ATNA can utilize node feature information effectively and optimize itself in the end-to-end network;(2)We proposed the unnormalized neighborhood convolutional layer(UNCL)based on the Gaussian mixture model(GMM).Firstly,UNCL learns a map based on the GMM.Through the map,a convolution parameter can be assigned to each node according to the relationship between neighborhoods and the center node.Then,like the CNN,UNCL takes convolution parameters as weights to calculate the eigenvalue weight sum of all nodes in the neighborhood field as a result of the convolution operation.UNCL can carry out convolution operation on non-normalized local neighborhood directly and avoid the loss of information caused by normalization.Based on ATNA and UNCL,we designed an unnormalized neighborhood graph convolutional network model with an attention mechanism(AT-UNGCN),which can solve the graph classification in an end-to-end fashion.In this paper,we compare our methods with the graph kernels and other mainstream graph convolutional neural networks through the experiments on the benchmark datasets and give the analysis of ATNA,UNCL and AT-UNGCN.The experimental results show that AT-UNGCN achieves good classification effect and shows excellent performance.Moreover,ATNA can utilize node feature information effectively and UNCL can avoid information loss caused by normalization,both of which can further improve the accuracy of graph classification.