SPGCL: Simple yet Powerful Graph Contrastive Learning via SVD-Guided Structural Perturbation

Graph Neural Networks (GNNs) are highly sensitive to structural noise, including spurious or missing edges caused by adversarial attacks or non-adversarial imperfections. Existing graph contrastive learning methods typically rely on either random perturbations (e.g., edge dropping) to ensure view diversity, or purely spectral augmentations (e.g., SVD) to preserve global structural priors. However, random perturbations are structure-agnostic and may inadvertently remove critical edges, while purely SVD-based views tend to be dense and lack sufficient diversity. More fundamentally, despite their complementary strengths, there exists no straightforward way to integrate these two fundamentally different paradigms: one operates on discrete edge removal while the other relies on continuous matrix factorization. To bridge this non-trivial gap, we propose SPGCL, a simple yet powerful framework for robust graph contrastive learning via SVD-guided structural perturbation. Our key insight lies in leveraging a recently developed SVD-based perturbation method, which generalizes the classical structural perturbation theory from symmetric to arbitrary graphs. We creatively adapt this technique to contrastive learning by designing a concise two-stage perturbation strategy: first, a lightweight stochastic edge removal injects necessary diversity; second, truncated SVD is applied to derive a structure-aware scoring matrix, followed by sparse top-$P$ edge recovery. This synergistic integration delivers three key advantages: (1) Robustness to accidental deletion - important edges removed in the first stage can be recovered by SVD-guided scoring, mitigating harmful effects while preserving diversity. (2) Enrichment with missing links - the SVD-based scoring can introduce new, semantically meaningful edges absent in the original graph, creating a more informative contrastive view. (3) Controllable structural discrepancy – by balancing edge removal and recovery rates, we explicitly control the edge count between views, ensuring that the contrastive signal stems from semantic structural differences rather than trivial edge-number gaps. Furthermore, we incorporate a contrastive fusion module regularized by a global similarity constraint to align node embeddings across views. Extensive experiments on ten benchmark datasets demonstrate that SPGCL consistently improves the robustness and accuracy of base GNNs, outperforming state-of-the-art graph contrastive learning and structure learning methods, thereby validating its simplicity, effectiveness, and innovative integration of previously disparate paradigms.