Abstract: MGAEs have gained significant attention due to their effectiveness in handling node classification tasks on graph-structured data.However,existing MGAE models face two main limitations during the pretraining of the encoder:semantic information loss,and similarity of embeddings for masked nodes.To mitigate these issues,this paper proposes a Contrastive Masked Graph Autoencoder model(CMGAE).Firstly,the masked graph and the original graph are separately fed into the online encoder and the target encoder to generate online embeddings and target embeddings,respectively.Then,an information supplementation module is employed to compare the similarity between the online embeddings and target embeddings,thereby recovering the lost semantic information.Simultaneously,the online embeddings are passed through a discriminator function and decoder.The discriminator function helps increase the variance of the embeddings for masked nodes,mitigating the issue of similar embeddings for masked nodes.The decoder reconstructs node features that are used to train the online encoder.Finally,the pretrained online encoder is utilized for node classification tasks.Node classification experiments are conducted on five transductive benchmark datasets and one inductive dataset.The results show that CMGAE achieves a transductive accuracy of 85.0%,73.6%,60.0%,50.5%,and 71.8% on the respective datasets,while the Micro-F1 score on the inductive dataset reaches 74.8%.These results demonstrate that CMGAE outperforms existing models.

Key words: Graph neural network, Node classification, Masked graph autoencoder, Graph self-supervised learning, Graph contrastive learning