Coarse-to-Fine Contrastive Learning on Graphs.

Area of Science:

• Graph Representation Learning
• Machine Learning
• Artificial Intelligence

Background:

• Contrastive learning (CL) is widely used for self-supervised node representation learning.
• Current CL methods rely on graph augmentation via structural or attribute perturbations.
• Existing methods overlook inherent information in graph augmentation, such as decreasing similarity and increasing node discrimination with perturbation degree.

Purpose of the Study:

• To propose a general ranking framework to incorporate prior information into contrastive learning.
• To leverage the ranking order among augmented graph views.
• To maintain discriminative information among nodes despite varying perturbation degrees.

Main Methods:

• Interpreting contrastive learning as a special case of learning to rank (L2R).
• Developing a general ranking framework to guide graph augmentation.
• Introducing a self-ranking paradigm for robust node discrimination.

Main Results:

• The proposed ranking framework effectively incorporates prior information into CL.
• The self-ranking paradigm maintains node discriminative information under varying perturbations.
• Experimental results demonstrate superior performance compared to existing supervised and unsupervised models on benchmark datasets.

Conclusions:

• The novel ranking framework enhances self-supervised graph representation learning.
• Incorporating prior information through ranking improves the quality of learned node representations.
• The method offers a more principled approach to graph augmentation in contrastive learning.