graph contrastive learning

To address the limited performance of graph contrastive learning (GCL) on heterophilic graphs, this paper proposes a lightweight, augmentation-free, and negative-sample-free dual-encoder contrastive framework. Methodologically, it employs a GCN encoder to capture structure-aware representations and an MLP encoder to suppress feature noise, naturally modeling raw node features and graph topology as complementary views; structural guidance is incorporated into the contrastive loss for end-to-end unsupervised training. Theoretical analysis shows that this design effectively mitigates feature confusion induced by heterophilous neighborhoods. Experiments demonstrate state-of-the-art performance on mainstream heterophilic graph benchmarks, with significantly lower computational and memory overhead than existing methods. Moreover, the framework exhibits superior scalability, controllable complexity, and enhanced adversarial robustness on homophilic graphs.

To address label sparsity and representation collapse in semi-supervised node classification, this paper proposes the Mixed Graph Contrastive Network (MGCN). MGCN jointly models discriminative structural information for both labeled and unlabeled nodes through latent-space interpolation augmentation and cross-view correlation constraints. Specifically, it introduces an interpolation-driven linear prediction consistency constraint to explicitly integrate limited supervised signals with self-supervised signals; concurrently, it enforces identity approximation of the cross-view correlation matrix to suppress representation redundancy and mitigate collapse. Evaluated on six benchmark datasets, MGCN consistently outperforms existing state-of-the-art methods. The implementation is publicly available, demonstrating strong effectiveness, robustness to varying label rates and graph perturbations, and generalization across diverse graph topologies and feature distributions.

This work addresses key challenges in unsupervised domain adaptation for graph classification—namely, label scarcity in the target domain, insufficient topological modeling, and substantial domain shift. To this end, we propose a dual-path coupled contrastive learning framework. Methodologically, we introduce the first integration of implicit graph convolutional networks (GCNs) and explicit hierarchical graph kernels (HGKs) into a two-branch architecture: the GCN branch captures local structural patterns, while the HGK branch encodes global semantic motifs. These branches are jointly optimized via multi-view coupled contrastive learning to achieve cross-domain semantic alignment and complementary representation enhancement. Crucially, the framework operates without any target-domain labels, mitigating both inadequate topological exploration and domain shift. Extensive experiments on multiple benchmark datasets demonstrate that our approach consistently outperforms state-of-the-art methods across diverse domain transfer settings, exhibiting superior robustness and generalization capability.

Noisy real-world graph data severely degrades the node classification performance of Graph Neural Networks (GNNs). To address this, we propose Low-Rank Graph Contrastive Learning (LR-GCL), a framework that jointly optimizes structural robustness—enforced via low-frequency priors—and semantic discriminability by integrating low-rank regularization into prototype-based contrastive learning. We theoretically derive the first tight generalization bound for graph contrastive learning, revealing how low-rank modeling fundamentally enhances both generalization and noise resilience. LR-GCL adopts a transductive learning paradigm and employs a linear classifier for efficient deployment. Extensive experiments on multiple benchmark datasets demonstrate an average accuracy improvement of 2.3% over state-of-the-art methods. Moreover, LR-GCL exhibits superior robustness under label noise and edge perturbations, validating its enhanced generalization capability and structural stability.

To address dimensionality collapse caused by feature aggregation in hyperbolic graph contrastive learning, this paper proposes the first hierarchical contrastive learning framework tailored for the Poincaré ball model. Methodologically, it formally defines uniformity requirements at both leaf-level and height-level hierarchies, and jointly optimizes a hierarchical alignment loss with an isotropic uniformity regularizer—constrained by a differentiable annular density penalty—to overcome limitations of Euclidean contrastive learning paradigms. The contributions are threefold: (1) theoretical modeling of hierarchical uniformity in hyperbolic space; (2) design of a differentiable annular density constraint to mitigate dimensionality collapse; and (3) consistent and significant performance gains across multiple hierarchical graph benchmarks, enhancing both feature space utilization and representation discriminability in downstream tasks.

Existing contrastive learning-based link prediction methods suffer from two major limitations: (i) a lack of theoretical grounding, and (ii) neglect of degree distribution imbalance and its adverse impact on contrastive optimization. This paper establishes the first theoretical analysis framework for contrastive learning in link prediction. We propose Edge-Balanced Augmentation (EBA), a degree-aware graph structural reweighting scheme that mitigates gradient bias in the contrastive loss between high- and low-degree nodes. Furthermore, we design a tailored contrastive loss function compatible with EBA and integrate it into an autoencoder architecture to enable end-to-end training. Extensive experiments on eight benchmark datasets demonstrate that our method consistently outperforms state-of-the-art models, validating its effectiveness, robustness, and generalizability.

Learning scalable node representations in graphs without node features and with only partial pairwise labels remains challenging. This work proposes Contrastive FUSE, a novel framework that, for the first time, integrates modularity-inspired structural learning with contrastive supervision, directly optimizing a spectral contrastive objective using community-aware structural signals and signed pairwise constraints. To enhance computational efficiency, the method introduces a lightweight gradient approximation to replace the costly exact modularity gradients and incorporates optimization techniques such as natural gradient decomposition and adaptive learning rate scaling. Experimental results demonstrate that Contrastive FUSE achieves classification performance on multiple benchmark graphs that is either superior or comparable to existing approaches—all without requiring node features—while significantly accelerating training compared to baseline methods.

Traditional self-supervised learning overly relies on data augmentation while neglecting semantic relationships among instances. To address this, we propose the first systematic incorporation of graph-structured modeling to explicitly capture inter-instance associations. Specifically, we construct a teacher–student dual-stream k-nearest neighbor (k-NN) graph and integrate graph neural networks (GNNs) to enable multi-hop message passing, thereby unifying local augmentations with global contextual information. Our approach introduces a k-NN-based dual-stream architecture coupled with a representation refinement mechanism, breaking away from the conventional paradigm that learns solely from intra-instance variations. Extensive experiments demonstrate consistent improvements in linear evaluation accuracy: +7.3% on CIFAR-10, +3.2% on ImageNet-100, and +1.0% on ImageNet-1K—outperforming state-of-the-art methods. These results validate the effectiveness and generalizability of explicitly modeling inter-instance relationships for self-supervised representation learning.

Existing graph contrastive learning methods primarily target undirected graphs and struggle to model the critical directional structures inherent in directed graphs. To address this, we propose S2-DiGCL—the first directed graph contrastive learning framework integrating dual perspectives in complex and real vector spaces. Its key contributions are: (1) complex-domain perturbation grounded in the magnetic Laplacian operator, coupled with adaptive edge-phase modulation, to explicitly encode directed adjacency relationships; and (2) a path-guided subgraph augmentation strategy that captures local asymmetry and higher-order topological dependencies. Evaluated on seven real-world directed graph benchmarks, S2-DiGCL achieves state-of-the-art performance—yielding average improvements of 4.41% in node classification and 4.34% in link prediction—while remaining compatible with both supervised and unsupervised settings.

Masked feature reconstruction (MFR) in graph self-supervised learning suffers from weak discriminability and conceptual disconnection from contrastive learning. Method: This paper theoretically establishes, under reasonable assumptions, the objective-function equivalence between MFR and node-level graph contrastive learning (GCL). Building on this insight, we propose Contrastive Masked Feature Reconstruction (CMFR)—a unified framework that introduces a novel contrastive reconstruction paradigm: original and reconstructed features serve as positive pairs, while masked nodes act as negatives. CMFR integrates a context-aware encoder and a customized negative sampling strategy. Contribution/Results: On multiple benchmark datasets, CMFR consistently outperforms GraphMAE and GraphMAE2, achieving up to 3.82% absolute improvement in both node and graph classification tasks, setting a new state-of-the-art in graph self-supervised learning.