Leveraging Structure–Attribute Mixed-View Augmentation for Improved Graph Contrastive Learning

Graph Contrastive Learning (GCL) offers a promising self-supervised paradigm for learning node representations, but existing methods typically rely on heuristic augmentations that perturb topology or attributes in isolation, often introducing semantic drift that degrades representation quality. We identify semantic preservation as a key but overlooked property of graph augmentation, and propose SAMA (Structure-Attribute Mixed-View Augmentation), a principled augmentation framework that jointly models structural connections and node attributes. SAMA integrates deep non-negative matrix factorization with spectral feature filtering to generate multiple semantically aligned augmented views, and adopts a two-stage training scheme of semantic-aware augmentation followed by multi-view contrastive learning. Extensive experiments on six benchmark datasets demonstrate that SAMA consistently outperforms thirteen state-of-the-art baselines in node classification, achieving improvements of up to 3.0 percentage points while maintaining strong performance under extreme label scarcity. These results highlight semantic-preserving augmentation as a representation learning insight for self-supervised graph learning.