Towards structure-aware data augmentation for high-degree graph neural networks

Traditional Graph Neural Networks (GNNs) underperform on High Degree Graphs (HDGs) due to over-smoothing, inefficiency, and structural semantic loss. To overcome these limitations, we propose SADA, a structure-aware data augmentation framework specifically designed for HDGs. Specifically, SADA integrates multi-scale structural embeddings by combining Node2Vec and Struc2Vec with an adaptive gating mechanism, effectively capturing both local and global information. Furthermore, it employs attention-guided graph sparsification using a learnable edge importance scoring function to remove redundant edges while preserving critical semantics. In addition, the adaptive GNN architecture incorporates degree-aware aggregation, a jump connection mechanism, and a personalized learning strategy to further enhance model expressiveness. Extensive experiments on eight real-world HDG datasets demonstrate that our approach achieves a 4.21% average improvement in accuracy and reduces training and inference time per epoch by approximately 10% compared to previous methods. The results validate our approach’s effectiveness and robustness. • Proposed a multi-scale embedding fusing Node2Vec and Struc2Vec via adaptive gating. • Proposed attention-guided graph sparsification using learnable edge scoring to optimize efficiency. • Built adaptive graph using degree-aware aggregation, jump connections, and personalized learning to boost expressivity.