FSGMI:Graph Node Classification based on Feature Selection and Global Mutual Information Technology

With the rapid development of Graph Neural Networks, node classification methods based on Graph Contrastive Learning have become a research focus. These methods learn node embeddings by maximizing the consistency of representations for the same node across different graph views. By generating multiple views of a graph for training, they enhance the model's robustness to noise and perturbations. However, existing methods still suffer from limitations in utilizing node attributes, mitigating embedding coupling, handling noisy features, and modeling mutual information. To address the above challenges, this paper proposes a model for graph node classification named FSGMI, which leverages feature selection and global mutual information techniques. The model incorporates a feature selection module to preprocess the graph, effectively reducing the interference of noisy nodes in the representation learning of other nodes. During contrastive learning, global mutual information is introduced into the loss calculation, and multiple discriminators are employed to score positive and negative samples collaboratively. This multi-discriminator design mitigates the risk of biased decisions caused by a single discriminator. Extensive experiments conducted on multiple public datasets demonstrate the effectiveness of the FSGMI model.