YANG Zhi-hao1，JIANG Wei-li1，DU Guo-dong2，XIANG Yan1，MA Lei1，SHAO Dang-guo1，YANG Jia-lin1
1（Faculty of Information Engineering and Automation，Kunming University of Science and Technology，Kunming 650504，China）2（Department of Artificial Intelligence，Xiamen University，Xiamen 361005，China）
Abstract：Existing works have shown that the Graph Convolutional Neural Networks have a poor understanding of the importance of nodes and cannot effectively use the information extracted fromeach convolutional layer.To settlethe issue，we propose an improved Graph Isomorphism Networks method for graph classification tasks.First of all，for the purpose of distinguishing the importance of nodes，we make nodes to get more weight that neighbor node has more node via degree matrix weighted convolution operators，it will give priority to the characteristics of these nodes when learning.Thus，in order to distinguish the importance of each layer of convolutional layers，we calculate the similarity of the features of the nodes of each layer，and weight the features of each layer of nodes viathe similarity.In the experiment，we compare our method withsix state-of-the-art neural network methods on four datasets.The experimental results have shown that our proposed multi-feature dynamic weighted graph convolutional network performs better than state-of-the-art neural network methods in the graph classification tasks.
2020, Vol. 41 
