Research On Graph Representation Learning Method Based On Contrastive Learning

Graph-structured data have appeared in a wide range of research fields and applications.Unlike text and images,the inherent complexity of graph-structured data has brought unprecedented challenges to machine learning.In recent years,graph neural networks have demonstrated powerful capabilities and are widely used in various graph data tasks.Although GNNs perform well on various tasks,most existing GNN models are shallow and can only visit neighbors within 2 or 3 hops due to oversmoothing.Therefore,only the labeled nodes and their neighbors effectively participate in the model optimization process,while the rich information in a large number of unlabeled nodes is ignored.In order to make full use of the information of unlabeled nodes,as an effective self-supervised learning method,contrastive learning technique is widely used in the field of graph data mining.The key idea of contrastive learning is to encourage the representation distance of similar data pairs to be close,and the representation distance of dissimilar sample pairs is pushed away.Because the true labels of all data samples cannot be accessed,for a given data point,its corresponding positive sample is usually obtained from the sample through data augmentation techniques,and other data points in the dataset are usually used as negative samples for this sample.Node-level graph contrastive learning methods basically do this.But this leads to two major problems,firstly,utilizing all nodes of a graph during contrastive learning can be very expensive,especially for large graphs.Second,many nodes that share the same label with the central node are used as negative samples,and then the contrastive learning strategy will separate the nodes with the same label(similar nodes),which will lead to poor performance.To address these issues,we first explore the feasibility of sampling only some nodes during graph contrastive learning.Unlike previous graph contrastive methods,we use jointly trained classification loss in contrastive learning to exploit as much supervisory signal as possible.Therefore,we propose a negative samples selection strategy that utilizes classification predictions to guide the selection of negative samples for sampling nodes.Then,we further combine this strategy for contrastive learning on graphs,and propose a framework named Graph Contrastive Learning with Negative Sample Selection Strategy.Then we proposed a graph comparison learning method based on feature grouping whitening,which achieves alignment by ensuring the consistency between "enhanced samples",but does not require the participation of negative samples in the contrastive learning process,but guarantees it through whitening operations.Homogeneity between samples.Because whitening has a "feature divergence" effect on samples,it avoids the representation of all samples from collapsing to a point,that is,dimension collapse.We demonstrate that our two proposed graph-contrastive learning methods can be trained quickly with less computational memory than current graph-contrastive learning methods,and they can demonstrate superior performance on semi-supervised node classification tasks on many datasets.