Toward Graph Self-Supervised Learning With Contrastive Adjusted Zooming.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph representation learning (GRL) is vital for analyzing graph data.
• Existing graph neural networks (GNNs) often require expensive labeled data.
• Unsupervised GRL methods face limitations like restricted contrastiveness and scalability.

Purpose of the Study:

• To introduce G-Zoom, a novel self-supervised GRL algorithm.
• To develop an adjusted zooming scheme for multi-scale signal extraction (node, neighborhood, subgraph).
• To enhance unsupervised GRL by overcoming limitations of existing methods.

Main Methods:

• G-Zoom employs a self-supervised approach using an adjusted zooming scheme.
• It generates two augmented graph views and applies contrastive learning across micro, meso, and macro scales.
• A parallel graph diffusion approach ensures scalability for large graphs.

Main Results:

• G-Zoom effectively extracts self-supervision signals from multiple graph scales.
• The adjusted zooming scheme allows for customizable viewpoints between node and subgraph levels.
• Extensive experiments show G-Zoom outperforms state-of-the-art methods on real-world datasets.

Conclusions:

• G-Zoom provides a powerful and scalable solution for unsupervised graph representation learning.
• The multi-scale contrastive approach enhances the understanding of graph data.
• This method offers a significant advancement over existing GRL techniques.