Beyond low-pass filtering on large-scale graphs via Adaptive Filtering Graph Neural Networks

Area of Science:

• Machine Learning
• Graph Theory
• Deep Learning

Background:

• Graph Neural Networks (GNNs) are vital for graph-structured data but face scalability challenges in industrial applications.
• Existing scalable GNNs often act as low-pass filters, losing crucial middle- and high-frequency information.
• This limits their effectiveness in capturing the full spectrum of data characteristics.

Purpose of the Study:

• To introduce Adaptive Filtering Graph Neural Networks (AFGNN), a novel GNN architecture.
• To enable the capture of all frequency information (low, middle, and high) on large-scale graphs.
• To address the scalability limitations and information loss issues of current GNN models.

Main Methods:

• AFGNN employs a two-stage process: pre-computed graph filters (low, middle, high-pass) for scalable feature extraction.
• A node-level attention mechanism creates customized filters per node, unlike uniform filters in spectral GNNs.
• The architecture supports mini-batch training for enhanced efficiency on large datasets.

Main Results:

• AFGNN successfully captures comprehensive frequency information from large-scale graphs.
• The model demonstrates superior scalability compared to existing scalable GNNs.
• AFGNN outperforms both spectral GNNs in frequency information capture and scalable GNNs in overall performance.

Conclusions:

• AFGNN offers a scalable solution for analyzing large-scale graph data by preserving all frequency information.
• The adaptive, node-level filtering mechanism provides customized feature extraction.
• AFGNN represents a significant advancement over existing GNNs for complex graph analysis tasks.