Design And Implementation Of Semi-supervised Node Classification Algorithm Based On Self-supervised Learning

In recent years,semi-supervised learning on graphs has been widely studied and applied,but the performance of semi-supervised learning on graph degrades severely when there are few labeled samples.Researches show that semi-supervised learning on graphs,especially contrastive learning on graphs,can mine the supervised information of graph data itself and learn more efficient models with less labeling costs.At present,the contrastive learning on graphs generates two views through graph data augmentation,and then use the graph neural network to encode the views to obtain the embedding representations of the node to calculate the contrastive loss.The representations from the same node are called positive pairs,and the representations from different nodes are called negative pairs,with the goal of maximizing the consistency between positive pairs.Although the contrastive learning on graphs has been widely studied,there are still problems such as selecting augmentation by tedious search and augmentation will inevitably change the graph semantic information.In addition,most contrastive learning on graphs treats all nodes except the anchor node as negative samples and embeds them far away from the anchor node,which seriously affects the effectiveness of contrastive learning on graphs.To solve the above problems,this paper proposes corresponding solutions based on graph contrastive learning.The main researches and results are as follows:(1)Aiming at the cumbersome and semantically corrupting problems of graph augmentation techniques,a semi-supervised node classification algorithm based on hierarchical contrastive learning is proposed.In this algorithm,the graph neural network is used as an encoder,and its different levels are used as views for contrastive learning on graphs.The algorithm uses semi-supervised contrastive loss on graph to provide rich supervised signals for learning.Experiments on the node classification task are conducted on benchmark datasets to verify the effectiveness of the proposed algorithm.(2)Aiming at the problem of poor robustness of negative samples selection,A contrastive learning on graph algorithm based on false negative samples elimination is proposed for semi-supervised node classification.The method first performs graph clustering,generates pseudo-labels according to the clustering results,and uses the information of pseudo-labels to identify the false negative example set.Finally,the false negative sample set is eliminated from the negative sample set during contrastive learning on graphs.Experimental results on citation network datasets show that the algorithm improves model performance.(3)Design and implement the graph semi-supervised node classification system based on self-supervised learning.The system includes the semi-supervised node classification algorithm based on hierarchical contrastive learning and other existing self-supervised learning algorithms,and implements the functions of data set pre-processing,algorithm operation,and result visualization.The system shows detailed descriptions of multiple algorithms and datasets in an intuitive and concise way,and provides references for read.In summary,this paper studies the applications of self-supervised learning on graph to semi-supervised learning on graph tasks.This paper proposes two graph contrastive learning methods to deal with the cumbersome and semantically corrupting problems of graph augmentation techniques,and the weak robustness of negative example selection.Based on this,a graph semi-supervised node classification system based on self-supervised learning is implemented to show the classification effects of multiple algorithms to users in an intuitive and easy-to-understand manner.