DEGD: Learning a Dual-Encoder Gated Decoder for graph data augmentation

This paper proposes a novel end-to-end graph data augmentation framework based on edge prediction and manipulation to enhance the generalization of graph neural network (GNN)-based models for node classification, addressing the limitations of existing methods in effectively balancing local and global feature aggregation under sparse and complex graph structures. We first introduce a Dual-Encoder Gated Decoder (DEGD) architecture that explicitly integrates local structure cues and global contextual information through two complementary components: a Structural Feature Fusion Encoder (SFFE) and a Stochastic Context-Aware Encoder (SCAE). The SFFE aggregates neighborhood evidence using complementary pooling heads to capture both dominant and subtle local patterns, while the SCAE samples multi-hop neighborhoods to model broader contextual dependencies. A learnable Gated Decoder then estimates pairwise gating weights and combines the adjacency scores derived from the two encoders into unified edge probabilities. To prevent the predictor from drifting away from the original graph, we further propose an adaptive dual-thresholding interpolation strategy that weights predicted and original adjacencies before sampling sparse graph variants. Finally, a GNN is employed to learn embeddings for node classification. Extensive experiments on ten benchmark datasets show that our approach achieves average micro-F1 improvements of 1.09% on CORA, 1.94% on CITESEER, 0.87% on PUBMED, 2.27% on PPI, 1.47% on AIR-USA, 2.63% on AMAZON COMPUTERS, 1.65% on AMAZON PHOTO, 2.43% on REDDIT, 1.89% on OGBN-ARXIV, and 1.65% on OGBN-PRODUCTS over state-of-the-art methods.