Toward Effective Graph Long-Tailed Learning With Noisy Labels.

Graph long-tailed learning has garnered significant research attention. However, prevailing works in this domain typically assume the cleanliness of training dataset labels, neglecting the reality of noisy labels in real-world data. Such challenges affect the generalization and robustness of graph neural networks (GNNs), making it crucial to address the learning problem with noisy labels in long-tailed data within graphs. While recent efforts have focused on the challenges of graph long-tailed learning or mitigating the impact of noisy labels individually, there is a conspicuous gap in concurrently addressing both issues. To fill this void, this article proposes a general graph learning framework, aiming to address this task and overcome existing limitations. In particular, we first introduce an enhanced robust representation learning to integrate local neighborhood and global discriminative information, which can effectively improve the robustness of node representations and the generalization of the model. Building upon this robust representation foundation, our framework further refines the focus on classifier optimization, which employs an effective calibration strategy and an adaptive clean sample selection mechanism, thus alleviating the graph long-tailed and noisy label issues. Extensive experiments on real-world datasets demonstrate that our approach outperforms strong baseline methods, highlighting its superior performance.