Augmentation-Free Graph Contrastive Learning of Invariant-Discriminative Representations

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Contrastive Learning (GCL) methods often rely on data augmentation for pretraining graph neural networks (GNNs).
• Current GCL approaches require extensive empirical tuning for data augmentation strategies and hyperparameters.
• This dependence on augmentation limits generalization and robustness.

Purpose of the Study:

• To propose an augmentation-free GCL method (iGCL) that learns invariant and discriminative representations.
• To eliminate the need for negative samples in GCL pretraining.
• To improve the generalization and robustness of GNNs.

Main Methods:

• Introduced invariant-discriminative loss (ID loss) for GCL.
• Minimized Mean Squared Error (MSE) between target and positive samples for invariance.
• Applied an orthonormal constraint to ensure representation discriminability and prevent collapse.
• Provided theoretical analysis linking ID loss to redundancy reduction, CCA, and IB principles.

Main Results:

• iGCL outperformed existing GCL baselines on five benchmark node classification datasets.
• Demonstrated superior performance across varying label ratios.
• Showcased significant robustness against graph attacks.
• Achieved excellent generalization capabilities.

Conclusions:

• iGCL offers an effective augmentation-free approach to GCL.
• The proposed ID loss successfully learns invariant and discriminative representations.
• iGCL significantly enhances the generalization and robustness of GNNs.