A learnable sampling method for scalable graph neural networks.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Handling large-scale graph data is a critical challenge in graph neural networks (GNNs).
• Current GNNs often rely on random sampling, which is detached from network propagation and not scalable to variable-weight models like graph attention networks.
• Existing sampling methods are often based on fixed statistical estimations, limiting their applicability to dynamic network architectures.

Purpose of the Study:

• To develop a learnable sampling method for GNNs that overcomes the limitations of traditional random sampling.
• To enable end-to-end training by integrating the sampling process dynamically with feature propagation.
• To create a scalable sampling solution applicable to various message-passing GNN models.

Main Methods:

• Proposed a novel learnable sampling method that allows for gradient calculation and unfixed probability sampling.
• Dynamically integrated the sampling process with the forward propagation of features within the neural network.
• Applied the learnable sampling method to GNNs, resulting in two new models designed for large-scale graph analysis.

Main Results:

• The learnable sampling method enables dynamic, probability-based node sampling, improving training efficiency.
• Achieved excellent accuracy on benchmark datasets with large graphs when combined with different GNN architectures.
• Demonstrated faster convergence rates and smaller loss values during training compared to previous methods.

Conclusions:

• The proposed learnable sampling method offers a scalable and effective solution for training GNNs on large graphs.
• This approach enhances model performance and training efficiency by dynamically integrating sampling with feature propagation.
• The method is generalizable to various message-passing GNN models, offering broad applicability in graph machine learning.