Multi-scale tree-guided contrastive learning for structure-aware graph representation

Existing graph contrastive learning (GCL) methods typically rely on random perturbations to generate augmented views. However, they often fail to explicitly model higher-order structural semantics. Most approaches apply uniform perturbation strengths across all nodes, ignoring local graph properties. This leads to limited structure awareness. In addition, prevailing contrastive objectives primarily focus on global alignment. As a result, they struggle to discriminate fine-grained topological variations. To address these limitations, we propose Multi-Scale Tree-Guided Contrastive Learning for Structure-Aware Graph Representation, MTGCL for brevity. MTGCL achieves a coordinated improvement in structural perception and discrimination capabilities from view construction to optimization objectives. Specifically, we first construct structural views based on multi-hop tree neighborhoods to explicitly encode multi-scale topological dependencies. Next, we introduce a dual-channel augmentation strategy that jointly perturbs both graph topology and node attributes. This generates diverse and structure-sensitive views. The combination ensures complementary information distortion—topology perturbation targets local structural roles, while attribute perturbation targets semantic content consistency. The selected fixed perturbation probabilities are validated through sensitivity analysis, which confirms that these parameters maintain a robust and stable performance regime, preventing either over-alignment or excessive structural destruction. Finally, we design a dual-path contrastive objective. It enforces semantic consistency across augmented views and structural discriminability for both structural and augmented views. This enables collaborative optimization of global and local representations. Experimental results on five widely used benchmark datasets demonstrate that MTGCL outperforms baseline methods, achieving an average improvement of 1.76%. Additional ablation and sensitivity analyses further validate the model’s robust generalization and parameter stability. These characteristics contribute to its explicit multi-order structural modeling.