Long-Tailed Graph Representation Learning via Dual Cost-Sensitive Graph Convolutional Network

Deep learning algorithms have seen a massive rise in popularity for remote sensing over the past few years. Recently, studies on applying deep techniques to graph data (e.g., public transport networks) been conducted. In node classification tasks, traditional neural network (GNN) models assume that different types of misclassifications an equal loss and thus seek maximize posterior probability sample nodes under labeled classes. The used realistic scenarios tend follow unbalanced long-tailed class distributions, where majority classes contain most vertices minority only small number nodes, making it difficult GNN accurately predict samples owing tendency this paper, we propose dual cost-sensitive convolutional (DCSGCN) model. DCSGCN is two-tower model containing two subnetworks compute misclassification cost. uses cost as ”complementary information” prediction correct perspective minimal risk. Furthermore, new method computing labels based topological information distribution. results extensive experiments demonstrate outperformed other competitive baselines real-world imbalanced graphs.