Heterogeneous graph neural network with relation-aware label propagation for unbalanced node classification

Node classification is one of the core downstream tasks of heterogeneous graph representation learning. However, existing heterogeneous graph neural networks (HGNNs) often exhibit bias toward the majority class, resulting in poor classification performance for the minority classes. Recently, some studies have begun to focus on the imbalance issue in homogeneous graphs. However, due to the inherent heterogeneity and imbalance of heterogeneous graphs, the exploration of imbalanced node classification in heterogeneous graphs remains under-explored. To bridge this gap, this paper investigates the representation learning on heterogeneous graphs and propose a novel model named Heterogeneous Graph Neural Network with Relation-aware Label Propagation (RLP-HGNN). To handle the heterogeneity, we design a relation-aware label propagation to obtain pseudo-labels of nodes in heterogeneous graphs. These pseudo-labels serve as a data augmentation strategy for subsequent phases. Different types of nodes may have different importance, and we adopt dual-level aggregation based on a type-attention mechanism for heterogeneous message passing among different relation subgraphs. To deal with the imbalance issue, we adopt different imbalance strategies to alleviate the classification bias in heterogeneous graphs, including Re-weight, Balanced Softmax, and PC Softmax. By combining relation-aware label propagation and dual-level aggregation into a multi-objective optimization problem, we train the whole model in an end-to-end fashion. We further study the performance of different methods under different imbalance ratio settings. With unbalanced strategies study, ablation analysis, and parameter sensitivity analysis, our experiments on heterogeneous graphs demonstrate the effectiveness and generalizability of our proposed approach in relieving imbalance issues.