Research On Graph Augmentation Based On Graph Contrastive Learning

In recent years,graph neural network has achieved great success in academia and industry because of its strong representation ability and high scalability.The current research on GNN model mainly focuses on the field of supervised learning that needs a large number of labels.However,there are the following problems in practical application:on the one hand,it is diffi-cult to obtain label information in large-scale graph due to security and labor cost;on the other hand,the complexity of graph structure data will make it more vulnerable to attack.There-fore,the GNN framework in the field of supervised learning has faced more and more severe challenges.It is urgent to propose an effective self supervised method on graph.As a self supervised learning method of graph,graph contrastive learning（GCL）was first inspired by the contrastive learning of computer vision,and soon became a more mainstream method with its remarkable performance.The general idea is to augment the graph data by adding random disturbance to obtain two different views of the same data,and then use con-trastive loss to maximize the consistency of the two views to train the model.Such a contrastive learning framework can make the last learned node embedding insensitive to the intentional augmentation disturbance,so as to improve its generalization.But nodes or edges on the graph have different sensitivity to noise due to their different importance.Therefore,the augmenta-tion strategy should give less noise to important places and more noise to unimportant places.However,most current GCL methods use the same noise distribution in all parts of the graph.In this dissertation,we attempt to introduce learnable graph augmentation into graph contrastive learning,and to utilizel₀regularization and other technologies to address these problems.The main contributions of this dissertation include the following two aspects:（1）This dissertation first designs a graph contrastive representation learning model with learnable augmentation based on drop factors（GCLA）.The model no longer allows each node or edge on the graph to use a unified or artificially adapted drop probability,but learns a dynamically changing drop probability.And we also usel₀regularization technology strengthens the sparsity of graph,so as to prevent too much mutual information from reducing the generalization of node embedding.Experiments on the node classification dataset show that the learnable augmentation method is helpful to improve the processing ability of the model,whether it is used in nodes,edges or features.（2）Based on the GCLA model,a graph contrastive representation learning model with learn-able augmentation based on graph structure information（GCLAGS）is further proposed.The model no longer uses the simple random initialization probability factor,but uses an-other independent graph convolution neural network to extract the information that can judge the importance of nodes or edges,so as to generate the probability factor more suit-able for the task.The model can use different graph convolution neural network extractors according to different characteristic graphs,so it has strong adaptability.A series of ex-periments on multiple node classification datasets show that GCLAGS can significantly improve the processing ability of the model.