This paper proposes a robust node classification method, RNCGLN, which simultaneously addresses graph noise and label noise in graph data. The method consists of three modules: a graph self-improvement module, a multi-head self-attention module, and a label self-improvement module. The graph self-improvement module uses a graph contrastive loss to learn local graph information and handle graph noise. The multi-head self-attention module captures global graph information to enhance node representation. The label self-improvement module constructs a classifier using node representations and original labels to detect and correct label noise. The method improves node classification performance by combining local and global graph learning, and by using pseudo graphs and pseudo labels to handle both types of noise. Experimental results show that RNCGLN outperforms existing methods in node classification under various noise conditions. The method is robust to both graph and label noise, and it effectively complements graph and label information to improve classification accuracy. The method is evaluated on four datasets and compared with several existing methods, demonstrating its effectiveness in handling noise in graph data.This paper proposes a robust node classification method, RNCGLN, which simultaneously addresses graph noise and label noise in graph data. The method consists of three modules: a graph self-improvement module, a multi-head self-attention module, and a label self-improvement module. The graph self-improvement module uses a graph contrastive loss to learn local graph information and handle graph noise. The multi-head self-attention module captures global graph information to enhance node representation. The label self-improvement module constructs a classifier using node representations and original labels to detect and correct label noise. The method improves node classification performance by combining local and global graph learning, and by using pseudo graphs and pseudo labels to handle both types of noise. Experimental results show that RNCGLN outperforms existing methods in node classification under various noise conditions. The method is robust to both graph and label noise, and it effectively complements graph and label information to improve classification accuracy. The method is evaluated on four datasets and compared with several existing methods, demonstrating its effectiveness in handling noise in graph data.