Research On Few-Shot Node Classification Method Based On Graph Neural Network

Currently,the way information is transmitted in society is constantly iterating,and the structure of data is becoming increasingly complex.In many fields such as bioinformatics,social network analysis,and transportation network planning,graphs are widely used due to their unique data structure and excellent representation capability.With the continuous progress of artificial intelligence and deep learning,graph representation learning is also constantly advancing.Among them,the node classification problem in graph neural networks is an important research direction,which can effectively represent non-Euclidean space graph structural data and has received increasing attention and research.However,in the real world,labeled sample data is often scarce and labeling costs are high,so few-shot node classification learning has gradually become a hot research direction.The core issue is how to train models using only a small amount of labeled data to have good generalization ability.Currently,meta-learning methods and data augmentation methods are two important research directions.However,traditional meta-learning-based technical models are often limited by data scarcity and are difficult to train well on meta-tasks,resulting in poor final performance.In addition,meta-learning-based methods have slow training speeds,and data augmentation methods usually only focus on labeled sample data,ignoring a large amount of unlabeled data and having insufficient representation ability for augmented samples.To address these problems,this paper proposes two improvement strategies:(1)Propose model: Meta-Learning Few-Shot Node Classification with Subgraph Sampling(SSM-FSNC).First,use two layers of GCN to perform embedding learning on all data to obtain feature representations.Next,use information entropy sorting to select the top K samples for each new class.Then,generate feature sample data for these samples according to standard deviation and variance,and adjust and correct them with base class data samples.Finally,use Gaussian distribution to generate new sample data from the newly generated feature data,and generate subgraphs for these new data.Each subgraph is treated as a meta-task,put into the MAML framework for learning,and fine-tuned on the query set for few-shot node classification.(2)Propose model: Data Augmentation Few-Shot Node Classification with Confidence Interval Label Propagation(CPD-FSNC).Although the subgraph sampling meta-learning model can greatly improve the performance of few-shot node classification,it is highly dependent on the representation ability of generated subgraphs,and the slow training speed of meta-learning has a significant impact on the final performance of the model.Therefore,this paper proposes graph data augmentation for few-shot node classification,which effectively alleviates the problem of too long training time for the first method,and then uses confidence label propagation to improve the feature representation ability of nodes.In summary,this paper proposes two algorithm models to solve the problems of few-shot node classification research,each of which improves the performance of the algorithm through different improvement strategies.Many experiments and analyses have verified the effectiveness of these methods.