SORAG: Synthetic Data Over-Sampling Strategy on Multi-Label Graphs

Abstract

In many real-world networks of interest in the field of remote sensing (e.g., public transport networks), nodes are associated with multiple labels, and node classes are imbalanced; that is, some classes have significantly fewer samples than others. However, the research problem of imbalanced multi-label graph node classification remains unexplored. This non-trivial task challenges the existing graph neural networks (GNNs) because the majority class can dominate the loss functions of GNNs and result in the overfitting of the majority class features and label correlations. On non-graph data, minority over-sampling methods (such as the synthetic minority over-sampling technique and its variants) have been demonstrated to be effective for the imbalanced data classification problem. This study proposes and validates a new hypothesis with unlabeled data over-sampling, which is meaningless for imbalanced non-graph data; however, feature propagation and topological interplay mechanisms between graph nodes can facilitate the representation learning of imbalanced graphs. Furthermore, we determine empirically that ensemble data synthesis through the creation of virtual minority samples in the central region of a minority and generation of virtual unlabeled samples in the boundary region between a minority and majority is the best practice for the imbalanced multi-label graph node classification task. Our proposed novel data over-sampling framework is evaluated using multiple real-world network datasets, and it outperforms diverse, strong benchmark models by a large margin.