Research On Few-shot Node Classification Methods Based On Graph Neural Network

Entities in the real world often interact and connect into complex graph networks,Node classification is a key task on graph networks,such as predicting users’ intentions on e-commerce networks,and inferring paper topics on citation networks.Therefore,the problem of node classification on the graph has aroused great interest of researchers.However,in the graph network of practical applications,the number of labeled nodes of all classes has a long tail distribution,that is,most classes have only a few labeled nodes.The traditional graph neural network model relies on the monitoring information of a large number of labeled nodes to train the classification model.This problem is the bottleneck to further improve the performance of the node classification task,Especially in the huge network with many nodes and sparse connections between nodes.Therefore,the few-shot node classification on graph networks is very challenging and practical.The few-shot node classification method aims to learn an effective model to classify graph nodes with few labels.The existing few-shot node classification method still uses the meta-learning paradigm to learn the model on a large-scale basic data set,and then applies it to new classes with only a few labeled samples on the target data set.The scarcity of labeled samples and the difference of feature distribution between sub-tasks will restrict the performance of Few-shot node classification.This paper systematically studies the few-shot node classification from these two different perspectives.The main research contents are as follows:(1)In order to effectively train the model and improve the robustness and reliability of the model under sparse labeled samples,a graph adaptive prototype networks(GAPN)model is proposed,which weights the importance of different support set instances to avoid the performance constraints of the classification results caused by the outliers of the sparse support set instances.Experiments on graph data sets show that the adaptive class prototype is effective for the classification performance of few shot nodes.(2)Aiming at the problem that the graph encoder cannot capture the feature distribution differences between subtasks,a few-shot node classification method based on the subtask feature re-representation(SFR)is proposed.This method uses the feature re-representation module to capture the difference of feature distribution between different tasks,and makes double corrections to the feature representation of scarce support set instances and query set instances,so as to capture the difference of important feature distribution between different tasks,which can effectively alleviate the impact of feature deviation on classification performance.Experiments on graph data sets show that the model correction module is effective for the classification performance of small sample nodes.By analyzing the problems existing in the existing few-shot node classification methods,this paper designs two few-shot node classification methods,which can effectively avoid the impact of the intra-class outliers of the rare support set instances and the feature distribution differences between sub-tasks on the classification performance,and provides two new methods and research ideas for dealing with the few-shot node classification problem.