Less is More: Towards Simple Graph Contrastive Learning

TL;DR

This work tackles the difficulty of graph contrastive learning on heterophilic graphs by advocating a minimal, augmentation-free approach that uses two complementary views: a GCN-based structural view and an MLP-based feature view. By ensuring weak correlation between the noises of these views, the model achieves strong noise cancellation and outperforms state-of-the-art methods on heterophilic benchmarks while remaining efficient and robust. The authors provide theoretical insights into noise decoupling and validate the method through extensive experiments, including adversarial robustness tests and complexity analyses. Overall, the paper argues that simplicity, via a two-view, augmentation-free design, can outperform more complex GCL pipelines in challenging graph regimes.

Abstract

Graph Contrastive Learning (GCL) has shown strong promise for unsupervised graph representation learning, yet its effectiveness on heterophilic graphs, where connected nodes often belong to different classes, remains limited. Most existing methods rely on complex augmentation schemes, intricate encoders, or negative sampling, which raises the question of whether such complexity is truly necessary in this challenging setting. In this work, we revisit the foundations of supervised and unsupervised learning on graphs and uncover a simple yet effective principle for GCL: mitigating node feature noise by aggregating it with structural features derived from the graph topology. This observation suggests that the original node features and the graph structure naturally provide two complementary views for contrastive learning. Building on this insight, we propose an embarrassingly simple GCL model that uses a GCN encoder to capture structural features and an MLP encoder to isolate node feature noise. Our design requires neither data augmentation nor negative sampling, yet achieves state-of-the-art results on heterophilic benchmarks with minimal computational and memory overhead, while also offering advantages in homophilic graphs in terms of complexity, scalability, and robustness. We provide theoretical justification for our approach and validate its effectiveness through extensive experiments, including robustness evaluations against both black-box and white-box adversarial attacks.

Figures (7)

• Figure 1: Performance–complexity trade-off of GCL methods on the Wisconsin and Roman datasets. Accuracy is plotted against training time (log scale, in seconds), with marker color indicating storage cost. Our GCN-MLP achieves the best performance with minimal complexity, while OOM cases are marked with red crosses. OOM refers to out of memory on an NVIDIA RTX A5000 GPU (24GB).
• Figure 2: For a sample node $v_i$, the noise $\bn_i$ (left panel) and $\bn^{(2)}_i$ (right panel) of a random selected node $v_i$ from the Cornell dataset. We see that $\bn^{(2)}_i$ displays a more random behavior.
• Figure 3: Performance comparison across feature dimension and GCN layers
• Figure 4: Performance comparison under adversarial attacks on Cora and Photo datasets
• Figure 5: Performance v.s. noise correlation on Cornell

Theorems & Definitions (3)

• proof : Proof of \ref{['lem:oba']}
• proof : Outline of the proof of \ref{['prop:gtf']}
• proof