Robust node classification against label distributional shift

Abstract

Node classification in graph data is an important problem in graph mining and machine learning, aiming is to predict the accurate label for each node in a graph. However, a major challenge in evaluating the performance of node classification algorithm lies in the label distribution shift between the training data and real data. The performance of node classification algorithm can be affected by label distribution shift, making it difficult to estimate the performance on real data from the training data performance. How can we train a node classification model that is robust to label distribution shift? Node embeddings obtained through graph neural networks have been widely utilized for various tasks, including node classification and link prediction. In this paper, we propose SCC, which uses self-supervised learning to train graph neural networks without label information and performs non-parametric node classification using the learned embeddings. SCC calculates the label probabilities of nodes directly from the node-specific embeddings obtained by self-supervised learning, without the need for additional parameters. This paper demonstrates that SCC exhibits (1) robustness: minimal performance degradation due to label distribution shift, (2) accuracy: achieving performance of 98-113% compared to non-linear node classification models, and (3) non-parametricity: conducting node classification without additional model parameters, as experimentally validated