Degree-Aware Graph Neural Network Quantization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph neural networks (GNNs) have shown great success in various tasks.
• Existing quantization methods for convolutional neural networks face challenges when applied to GNNs.
• These challenges include inflexible parameter scaling and uneven responses due to varying node degrees.

Purpose of the Study:

• To develop an effective quantization strategy for GNNs.
• To overcome the limitations of fixed-scale parameters and node degree variations in GNN quantization.
• To enhance the accuracy and applicability of quantized GNNs.

Main Methods:

• Introduced learnable scale parameters optimized jointly with GNNs.
• Proposed degree-aware normalization to handle diverse node degrees.
• Conducted experiments on various tasks, baselines, and datasets.

Main Results:

• The proposed method demonstrates superior performance compared to existing state-of-the-art quantization techniques.
• Learnable scale parameters provide flexibility across different GNN architectures and tasks.
• Degree-aware normalization effectively addresses accuracy degradation caused by node degree variations.

Conclusions:

• The developed quantization approach significantly improves GNN performance.
• The method offers a more robust and adaptable solution for GNN quantization.
• This work paves the way for efficient deployment of GNNs in resource-constrained environments.