Single-View Graph Contrastive Learning with Soft Neighborhood Awareness

TL;DR

The paper addresses the reliance on cross-view augmentations in graph contrastive learning by proposing SIGNA, a single-view framework that employs soft neighborhood awareness through a dropout-based encoder, stochastic neighbor masking, and the Norm-JSD objective. Under Norm-JSD, the expected neighbor similarity is $1-\alpha$ and non-neighbor similarity is $0$, enabling a controlled, discriminative embedding space. Empirically, SIGNA outperforms state-of-the-art unsupervised baselines across transductive and inductive node classification and clustering, with up to 21.74% gains on PPI and substantial inference-time speedups when using an MLP encoder. The work demonstrates that carefully moderated neighborhood signals can match or surpass cross-view CL while reducing augmentation design complexity and computation, offering practical benefits for scalable graph representation learning.

Abstract

Most graph contrastive learning (GCL) methods heavily rely on cross-view contrast, thus facing several concomitant challenges, such as the complexity of designing effective augmentations, the potential for information loss between views, and increased computational costs. To mitigate reliance on cross-view contrasts, we propose \ttt{SIGNA}, a novel single-view graph contrastive learning framework. Regarding the inconsistency between structural connection and semantic similarity of neighborhoods, we resort to soft neighborhood awareness for GCL. Specifically, we leverage dropout to obtain structurally-related yet randomly-noised embedding pairs for neighbors, which serve as potential positive samples. At each epoch, the role of partial neighbors is switched from positive to negative, leading to probabilistic neighborhood contrastive learning effect. Furthermore, we propose a normalized Jensen-Shannon divergence estimator for a better effect of contrastive learning. Surprisingly, experiments on diverse node-level tasks demonstrate that our simple single-view GCL framework consistently outperforms existing methods by margins of up to 21.74% (PPI). In particular, with soft neighborhood awareness, SIGNA can adopt MLPs instead of complicated GCNs as the encoder to generate representations in transductive learning tasks, thus speeding up its inference process by 109 times to 331 times. The source code is available at https://github.com/sunisfighting/SIGNA.

Figures (8)

• Figure 1: (a) Augmentation based cross-view contrast; (b) Non-augmentation based cross-view contrast; (c) Input-latent cross-view contrast; (d) Single-view contrast (ours).
• Figure 2: Global and local homophily statistics. Top: global homophily ratios on different datasets. Bottom Left: the distribution of local homophily counts on Photo. Bottom Right: the distribution of local homophily ratios on Photo.
• Figure 3: The framework of our proposed SIGNA. The contrast is conducted within a single graph view, and thus SIGNA has only one branch. The encoder with dropout implicitly provides more embedding combinations for robust contrast. The role of neighbors is variable, while non-neighbors are fixed as negative samples. The normalized JSD estimator facilitates better contrastive effect. The goal of SIGNA is to realize soft neighborhood awareness.
• Figure 4: The soft neighborhood awareness of SIGNA yields better performance.
• Figure 5: Representation visualization via t-SNE.

Theorems & Definitions (6)

• Definition 3.1: Global Homophily Ratio
• Definition 3.2: Local Homophily Count
• Definition 3.3: Local Homophily Ratio
• Theorem 4.1: Probabilistic Neighborhood Contrastive Learning
• proof
• Corollary 4.1