Study In Theories And Methods Of Graph Classification Based On Graph Neural Networks

A large number of complex networks exist in real scenarios,such as citation networks,social networks,protein molecular networks,etc.Therefore,mining the inherent information of complex networks has high practical value.Modeling the relationship between entities in a network based on graph algorithms can effectively mine the information of the network,which has been studied a lot and has obtained superior results.A complex network can be abstracted into a graph.The purpose of graph embedding is to learn a real-valued and low-dimensional vector for each entity in a graph,which can well preserve the topological structure of the graph and the content information of entities.The vector representation of the entire graph can be further obtained by some kind of operator,which can be applied to subsequent graph-level analysis tasks.Traditional graph embedding methods cannot reflect the inherent relationship between nodes in a graph due to the complexity of graph data,and can not well preserve the graph topology,which will affect the effect of downstream network analysis tasks.Graph neural networks can be trained in the end-to-end fashion and driven by data and tasks.These kind of models can be directly applied to graph data to achieve efficient modeling on graph data,which makes them widely concerned in the task of mining graph data information.This paper explores graph representation learning and proposes different models for graph analysis tasks in three scenarios,including supervised graph classification,semi-supervised graph classification and few-shot graph classification.The works of this paper include the following two parts:(1)At present,most graph neural network models designed for the traditional graph classification task are inherently flat and do not learn the hierarchical representation of a graph.This limitation will not solve the task of graph classification effectively.To solve this problem,this paper proposes a graph neural network model based on hierarchical coarsening,which can extract the hierarchical structure information in a graph.Moreover,in order to utilize the rich information in the graph and better retain the original node information,a link loss is introduced to learn the node embedding for assisting in completing the task of graph classification.Experimental results show that the method can better retain the original topological structure of a graph and make the graph representation more robust.On several benchmark graph classification datasets,classification accuracies of the model on graph classification and semi-supervised graph classification tasks outperform most baselines.(2)Algorithms for solving the traditional graph classification task often ignore the fact that graph data are often lack of labels in practical applications.To solve this problem,this paper will study few-show graph classification.When solving the few-shot problem,the data whose label space does not intersect with that of the data to be classified are allowed to use for assisting in training a transferable model.To improve the transferable ability of a model,this paper proposes a graph neural network framework based on co-training.The framework can extract two sufficient and compatible views of data,thereby obtaining better generalization ability.Proposed framework consists of two parallel branches,each of which consists of a graph neural network architecture,an aggregation layer and a classifier.The initial graph embedding and data relation graph are extracted by the graph neural network architecture,and then combined by downstream aggregation layer to learn refined graph representation.The distributed graph data whose label space does not overlap with that of the target data are used to train the proposed model.Moreover,model parameters are updated under the federated learning settings for protecting data privacy.Experimental results demonstrate that proposed method is superior to all baselines on few-show graph classification task.