Research On Node Classification And Graph Classification Methods Based On Small Sample Learnin

With the widespread application of large-scale graph data,efficient and accurate analysis of graph data has become a research hotspot.Node classification and graph classification are two important tasks in graph data analysis.Although graph neural networks have made some progress in graph data analysis tasks,they require a large amount of labeled data for training.In practical application scenarios,labeled data is often scarce,with only one or a few labeled nodes and graphs as supervised objects,which poses a great challenge to classification tasks.In addition,since graphs belong to non-Euclidean space and have complex topological structures,traditional few-shot learning methods cannot be directly transferred to the field of graph data.Therefore,it has become an important issue to find appropriate few-shot learning methods for graph data analysis in current research.The paper investigates node and graph classification methods based on few-shot learning and graph neural networks.The specific work is summarized as follows:(1)To address the issue of weakened generalization performance of few-shot learners caused by feature distribution differences in node classification,this paper proposes an attribute-enhanced graph meta-learning method(AMGM).AMGM introduces a collaborative attention strategy in the node encoding process to enhance node attributes locally and globally and thus reduces the differences in node feature distributions.This method follows the scenario training mechanism,and the entire learning process is divided into two stages: meta-training and meta-testing.In the metatraining phase,AMGM uses gradient-based optimization algorithms to update network parameters.In the meta-testing phase,based on the network parameters learned in the meta-training phase,AMGM can quickly adapt to new node classification tasks.(2)To address the problem of neglecting the differences and interdependence between nodes in few-shot models,this paper proposes a graph relation network method,(GRN).Through a graph convolutional network,information is gradually transmitted between nodes,and a node importance evaluator is designed to calculate the importance score of each node that has message passing,to more accurately determine the amount of information carried by each node.GRN constructs a learnable metric module to compare the similarity between nodes more accurately.By inputting the node embedding of each task into the graph neural network,the output similarity score is used to determine the node category.(3)To address the problem of sparse graph labels leading to overfitting,this paper proposes an enhanced prototype network method,PGC(Prototype network-based Graph Classifier).Considering the dependency between different hierarchy levels,PGC introduces a hierarchical attention graph encoder to extract features from graphs.PGC uses a prototype network to construct a prototype for each category and use a measurement function to calculate the distance between each graph and prototype to enable efficient graph classification.Additionally,the method proposes an embeddingenhanced strategy to recombine graph features,enabling the model to learn a wider range of graph characteristics and improve its generalization ability.(4)Based on the two proposed node classification methods and one graph classification method,this paper designed and implemented a prototype system that includes node category prediction,node information analysis,and graph category prediction.