PI-GNN: Towards Robust Semi-Supervised Node Classification against Noisy Labels

Xuefeng Du, Tian Bian, Yu Rong, Bo Han, Tongliang Liu, Tingyang Xu, Wenbing Huang, Junzhou Huang

Abstract: Semi-supervised node classification on graphs is a fundamental problem in graph mining that uses a small set of labeled nodes and many unlabeled nodes for training, so that its performance is quite sensitive to the quality of the node labels. However, it is expensive to maintain the label quality for real-world graph datasets, which presents huge challenges for the learning algorithm to keep a good generalization ability. In this paper, we propose a novel robust learning objective dubbed pairwise interactions (PI) for the model, such as Graph Neural Network (GNN) to combat against noisy labels. Unlike classic robust training approaches that operate on the pointwise interactions between node and class label pairs, PI explicitly forces the embeddings for node pairs that hold a positive PI label to be close to each other, which can be applied to both labeled and unlabeled nodes. We design several instantiations for the PI labels based on the graph structure as well as node class labels, and further propose a new uncertainty-aware training technique to mitigate the negative effect of the sub-optimal PI labels. Extensive experiments on different datasets and GNN architectures demonstrate the effectiveness of PI, which also brings a promising improvement over the state-of-the-art methods.