Bringing Your Own View: Graph Contrastive Learning without Prefabricated Data Augmentations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Self-supervision is advancing graph learning, but current methods like GraphCL rely on manual data augmentation.
• This reliance on domain knowledge and trial-and-error limits the efficiency and scalability of graph representation learning.

Purpose of the Study:

• To develop an automated framework for graph contrastive learning (GraphCL).
• To address the limitations of manual data augmentation in GraphCL by learning a continuous prior in the parameter space of graph generators.
• To integrate principles of information minimization (InfoMin) and information bottleneck (InfoBN) for regularizing learned priors and preventing trivial solutions.

Main Methods:

• Extended GraphCL by moving from discrete augmentations to a learnable continuous prior in graph generator parameters.
• Leveraged InfoMin and InfoBN principles to regularize learned priors.
• Developed a bi-level optimization framework integrating contrastive learning, InfoMin, and InfoBN.

Main Results:

• The automated approach demonstrates competitive performance against state-of-the-art self-supervision methods on small graph benchmarks.
• Achieved superior generalizability on large-scale graphs compared to existing methods.
• Eliminated the need for human expertise or downstream validation in the augmentation process.

Conclusions:

• The proposed principled and automated framework significantly advances GraphCL by enabling learnable priors.
• This method offers a more efficient and generalizable approach to self-supervised graph representation learning.
• The successful application on diverse benchmarks highlights the potential of automated graph augmentation.