Self-Reinforced Graph Contrastive Learning

TL;DR

SRGCL (Self-Reinforced Graph Contrastive Learning), a novel framework that leverages the model's own encoder to dynamically evaluate and select high-quality positive pairs, is proposed, underscoring its adaptability and efficacy across various domains.

Abstract

Graphs serve as versatile data structures in numerous real-world domains-including social networks, molecular biology, and knowledge graphs-by capturing intricate relational information among entities. Among graph-based learning techniques, Graph Contrastive Learning (GCL) has gained significant attention for its ability to derive robust, self-supervised graph representations through the contrasting of positive and negative sample pairs. However, a critical challenge lies in ensuring high-quality positive pairs so that the intrinsic semantic and structural properties of the original graph are preserved rather than distorted. To address this issue, we propose SRGCL (Self-Reinforced Graph Contrastive Learning), a novel framework that leverages the model's own encoder to dynamically evaluate and select high-quality positive pairs. We designed a unified positive pair generator employing multiple augmentation strategies, and a selector guided by the manifold hypothesis to maintain the underlying geometry of the latent space. By adopting a probabilistic mechanism for selecting positive pairs, SRGCL iteratively refines its assessment of pair quality as the encoder's representational power improves. Extensive experiments on diverse graph-level classification tasks demonstrate that SRGCL, as a plug-in module, consistently outperforms state-of-the-art GCL methods, underscoring its adaptability and efficacy across various domains.

Figures (7)

• Figure 1: The core idea of SRGCL is that high-quality positive pairs enhance the model’s capability, enabling it to more effectively discriminate the quality of positive pairs.
• Figure 2: The proposed framework of SRGCL, which is the plug-in module for existing GCL frameworks. SRGCL comprises: (1) UPPG for generating positive pair candidate set $\mathcal{C}_{i}$, whose size is $c$, (2) GNN shared encoder $e(\cdot)$ for encoding the input graph and candidate positive pairs to vector $z_i, \tilde{z}_{i}^{c}$, and (3) MiPPS for selecting the final positive set $\mathcal{P}_{i}$ from $\mathcal{C}_{i}$.
• Figure 3: Performance versus different distance functions across different datasets on Graph_SR (a) AutoGCL_SR (b). The accuracy is normalized to L2-norm configuration. The black vertical line is the standard deviation of each configuration.
• Figure 4: Performance versus temperature constant $s$ on GraphCL_SR (a) and AutoGCL_SR (b). The light blue tunnel part is the standard deviation of SRGCL run in 5 times. The orange dashed line is the average accuracy of SRGCL without probabilistic optimization. Once $s=0$, the SRGCL degrades to its original design, which can be considered as the fully random implementation.
• Figure 5: Performance versus augmentation combination across 8 datasets on GraphCL_SR (p). The accuracy is normalized to N+E+A. The black vertical line is the standard deviation of each configuration.