Unified Robust Training for Graph Neural Networks Against Label Noise

Graph neural networks (GNNs) have achieved state-of-the-art performance for node classification on graphs. The vast majority of existing works assume that genuine labels are always provided training. However, there has been very little research effort how to improve the robustness GNNs in presence label noise. Learning with noise primarily studied context image classification, but these techniques cannot be directly applied graph-structured data, due two major challenges—label sparsity and dependency—faced by learning In this paper, we propose a new framework, UnionNET, noisy graphs under semi-supervised setting. Our approach provides unified solution robustly training performing correction simultaneously. key idea is perform aggregation estimate node-level class probability distributions, which used guide sample reweighting correction. Compared works, UnionNET appealing advantages. First, it requires no extra clean supervision, or explicit estimation transition matrix. Second, framework proposed train an end-to-end manner. Experimental results show our approach: (1) effective improving model against different types levels noise; (2) yields significant improvements over baselines.