Multi-Scale Subgraph Contrastive Learning

TL;DR

This work reveals that graph semantic information after augmentation is scale-dependent and not always aligned across views. It introduces MSSGCL, which generates global and local subgraph views, learns corresponding graph-level embeddings, and optimizes a three-way contrastive loss that aligns global-global and global-local relationships while enforcing dissimilarity between local-local views via a learnable regressor. The method achieves state-of-the-art results on eight real-world datasets in both unsupervised and semi-supervised graph classification, demonstrating the benefit of explicitly modeling multi-scale semantics. The findings underscore the importance of scale-aware augmentation and contrastive signals for robust graph representations with practical impact in chemistry, biology, and network analysis.

Abstract

Graph-level contrastive learning, aiming to learn the representations for each graph by contrasting two augmented graphs, has attracted considerable attention. Previous studies usually simply assume that a graph and its augmented graph as a positive pair, otherwise as a negative pair. However, it is well known that graph structure is always complex and multi-scale, which gives rise to a fundamental question: after graph augmentation, will the previous assumption still hold in reality? By an experimental analysis, we discover the semantic information of an augmented graph structure may be not consistent as original graph structure, and whether two augmented graphs are positive or negative pairs is highly related with the multi-scale structures. Based on this finding, we propose a multi-scale subgraph contrastive learning architecture which is able to characterize the fine-grained semantic information. Specifically, we generate global and local views at different scales based on subgraph sampling, and construct multiple contrastive relationships according to their semantic associations to provide richer self-supervised signals. Extensive experiments and parametric analyzes on eight graph classification real-world datasets well demonstrate the effectiveness of the proposed method.

Figures (3)

• Figure 1: Semantic similarity of subgraphs of different sizes.
• Figure 2: The overall architecture of MSSGCL (left). The original graph generates global views and local views pairs through random walks with controlled number of nodes, which are then fed into the encoder to obtain global and local representations. We maximize the similarity between global views and the similarity between global and local views by optimizing $\mathcal{L}_{gg}$ and $\mathcal{L}_{gl}$. The dissimilarity between local views is encouraged by optimizing the output of a learned similarity measure $\mathcal{L}_{ll}$. $f_{\theta _e}$ is the GNN-based encoder, which includes a backbone network followed by a multi-layer perceptron. $f_{\theta _d}$ (right) is a learnable regressor to measure the similarity between local views.
• Figure 3: Classification accuracies of MSSGCL on NCI1 under different hyper-parameters.