Simple yet Effective Diffusion-based Graph Data Augmentation via Complementary Diffusion Transfer

Graph Neural Networks (GNNs) have emerged as powerful tools for modeling relational data, demonstrating impressive performance across many downstream tasks. Unfortunately, their effectiveness is often hindered by inherent graph imperfections, such as noise and incompleteness. To enhance the robustness of GNNs, numerous graph data augmentation (GDA) techniques have been proposed, with diffusion-based methods receiving significant theoretical and empirical backing. However, existing diffusion-based GDA approaches predominantly rely on hard masking or truncation operations, which may inadvertently discard valuable propagated information, thereby limiting performance. To address this dilemma, we propose CoDiT (Complementary Diffusion Transfer), a novel approach that constructs two complementary diffusion views and compensates for missing diffusion signals through cross-view transfer. A key advantage of CoDiT lies in its lightweight design, as it introduces no additional generative modules and maintains high representation quality. Specifically, CoDiT decomposes a truncated diffusion matrix into two disjoint components, each forming a complementary view. In each view, CoDiT performs restart-based iterative diffusion using a view-specific masked propagation operator, while simultaneously transferring information from the complementary view to enrich the current one. Original and transferred representations are fused via an interpolation operation, and a shared classifier is trained under a hybrid objective that combines supervised loss and consistency regularization loss. Extensive experiments on five real-world datasets demonstrate the superiority of the proposed CoDiT over thirteen competitive baselines.