The Study Of Negative Sampling Method In Link Prediction With Graph Neural Network

As the outstanding performance of the graph neural network(GNN)models,numerous studies have been attracted to apply them in link prediction.In these studies,the observed edges in a network are utilized as positive samples,and unobserved edges are randomly selected as negative samples.However,there are problems in randomly selecting unobserved edges as negative samples.First,some edges actually exist in the network but are not observed.Second,some unobserved edges can be easily distinguished from the positive samples.Using these two kinds of unobserved edges as negative samples results in false and easy negative samples,which hinder the performance of GNN models.To address these problems,the main work in this study is listed as follows:(1)To reduce the impact of false and easy negative samples,a curriculum learningbased negative sampling method is proposed in this study.In this method,the difficulty of distinguishing unobserved edges from positive samples is evaluated.In the negative sampling range,the difficulty of unobserved edges increases gradually.This method not only reduces the possibility of generating false negative samples,but also gradually increases the difficulty of negative samples to promote the performance of GNN models.(2)To further improve the performance of GNN models in link prediction,a policy gradient-based negative sampling method is designed in this study.In this method,a negative sample selector is designed to automatically select unobserved edges as negative samples,which takes a policy network as its core module.In the selecting process,the negative sample selector obtains the state of the unobserved edges as the input of the policy network and determines whether to select these edges as negative samples.The two negative sampling methods designed in this study can reduce the impact of false and easy negative samples in the training process of GNN models.Experimental results show that the performance of these two methods can exceed three commonly used negative sampling methods in most cases.Among them,the policy gradient-based negative sampling method achieves the best performance in most experimental results.