| Few-shot learning(FSL)has widely been proposed to train machine learning models(e.g.,Convolutional Neural Network(CNN))with generalization ability on limited supervised information,which has been well discussed on Euclidean data.However,in real applications,non-Euclidean data has shown explosive growth.Therefore,the research in FSL on non-Euclidean data is also particularly important.The most representative non-Euclidean data is graph data,and there have been many research works that apply FSL to graph data.The current popular research direction on graph data is few-shot node classification(FSNC),which aims to predict unlabeled nodes on graphs using limited labels.Previous FSNC models were mainly divided into two categories,i.e.,meta-learning-based models and metric learning-based models.However,they both overly focused on mining prior knowledge of the base classes and did not consider the rich information of unlabeled nodes in the novel classes,resulting in insufficient utilization of information.Self-training algorithms can mine rich information in unlabeled nodes by assigning pseudo-labels to high-confidence unlabeled nodes.Therefore,this paper proposes two FSNC models based on self-training algorithm to address the issues in meta-learning and metric learning.(1)A self-training FSNC with information augmentation.The model introduces a selftraining algorithm to select high-confidence unlabeled nodes and assign them with pseudolabels,achieving first-round data augmentation and mining rich information from unlabeled nodes in the novel classes.Then,the support set and pseudo-label set in the novel classes are jointly fed into a mapping function(e.g.,a Multi-Layer Perceptron(MLP))to generate a new node set,which is used to update the model and achieve second-round data augmentation.Throughout the entire process,this paper employ self-training algorithm as a data augmentation intermediary,resulting in two rounds of data augmentation,termed as information augmentation.(2)A self-training FSNC with knowledge distillation.Although good FSNC models can be trained using information augmentation,the key to making information augmentation work is the self-training module.However,two issues are prone to arise when implementing the self-training module.First,self-training module can mislead the training direction of the model when incorrect pseudo-labels are used.Second,self-training module can lead to model over-fitting when overconfident pseudo-labels are used.These two issues illustrate that the quality of pseudo-labels affects the performance of self-training module.Therefore,this paper proposes to improve the quality of pseudo-labels through knowledge distillation,thereby enhancing the performance of self-training module.Two self-training models are proposed in this paper to address the issue that previous FSNC models did not consider the rich information in the novel classes.Experimental results demonstrate the effectiveness of the proposed models. |
