Structure-based Updatable Graph Pooling For Graph Classification

In recent years,deep neural networks have developed rapidly and play an important role in face detection and speech recognition.When stepping into big data era,more and more data can be modeled as graph relationship,such as compound molecules,social networks.Each node in the graph contains its features and the relationship with other nodes,and the graph neural network can utilize the information to complete tasks such as node classification,graph convolution,link prediction and so on.However,most of the existing graph pooling models applied to graph classification are used to compute the importance scores of nodes through complex measurement methods,which correspondingly leads to the increase of the computational complexity.In addition,sampling nodes according to a certain strategy will inevitably lose part of the information of unsampled nodes,which will lead to incomplete graph representation.The purpose of this research is to design a novel pooling model and form a complete graph representation under the condition of considering the above factors.In this paper,a novel graph pooling model is proposed.In the pooling process,it takes into account both the graph structure and the features of nodes in the pooling process,which maximizes utilization of graph information theoretically.In addition,in order to retain the information of unsampled nodes,a novel strategy for updating these nodes is proposed.The information of each node in the graph is updated through the strategy,and the information of the original graph is retained as much as possible.The model is named Structure-based Updatable Graph Pooling,referred to as SUGPool.In the training process of the model,the size of the parameters matches the dimensions of the nodes' features,so it will not increase with the increase of the size of the input graph,which greatly speeds up the training process of the model and reduces the complexity of space and time.Compared with other graph pooling methods on seven graph datasets,the results show that the graph pooling model designed in this paper can generate better graph representation.Among seven datasets,we achieve the best performance on six datasets,which is a considerable improvement.