Graph topology adaptive judgment against node label noise

Graph Neural Networks (GNNs) have exhibited remarkable capabilities in processing graph data. Nevertheless, their performance is highly dependent on the labeled data, making them vulnerable to label noise. Existing methods often improve the robustness of node classification by adding trusted edges to the graph. However, most of them overlook the impact of potential noisy edges on the model's robustness, leading to a notable decline in performance as the average degree of the dataset increases. In this paper, we provide a theoretical explanation for the performance degradation observed in existing methods on datasets with high average degrees. Building on this insight, we propose the Graph Topology Adaptive(GTA) model, which incorporates EdgeBoost Module to add trusted edges based on node latent space similarity and EdgePrune Module to eliminate untrusted edges through optimized screening mechanism. Two modules work collaboratively to adaptively adjust the graph topology and generate final predictions. Both theoretical analysis and extensive experimental results validate the effectiveness of the GTA model.