SGB-Net: Scalable Graph Broad Network.

Area of Science:

• Graph learning
• Machine learning
• Data science

Background:

• Real-world graph data is complex and self-evolving, posing challenges for current graph learning methods.
• Existing methods struggle with representation learning limitations and require complete retraining for evolving graphs, especially without new labels.

Purpose of the Study:

• To propose a scalable graph broad network (SGB-Net) that addresses the limitations of current graph learning methods.
• To enhance graph embedding and enable scalability for evolving graph data.

Main Methods:

• Introduced the graph feature broad transformation (GFBT) layer to expand graph feature space and build models broadly.
• Developed two update algorithms: SGB-Net-U for label-free incremental learning and SGB-Net-S for label-assisted incremental learning.
• Designed SGB-Net to embed graphs discriminatively across various scales and adapt to graph expansion without retraining.

Main Results:

• SGB-Net enhances graph embeddings by constructing expandable feature spaces.
• SGB-Net-U leverages unsupervised knowledge for label-free graph incremental learning.
• SGB-Net-S provides scalability in traditional incremental learning scenarios.
• Experiments on 15 benchmark datasets show SGB-Net outperforms state-of-the-art Graph Neural Networks (GNNs) in effectiveness and scalability.

Conclusions:

• SGB-Net offers a scalable framework for graph learning that enhances representation and adapts to evolving graphs.
• The proposed GFBT layer and update algorithms enable improved performance without complete retraining.
• SGB-Net demonstrates superior effectiveness and scalability compared to existing GNNs.