Graph convolutional network with adaptive grouping aggregation strategy.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Graph convolutional networks (GCNs) face performance bottlenecks due to naive node aggregation functions, limiting their theoretical expressivity and practical application.
• Existing learning-based aggregation strategies lack focus on expressivity and standardized experimental evaluation.
• Naive aggregation functions fail to retain sufficient node information, resulting in less discriminative features and a performance gap.

Purpose of the Study:

• To address the limitations of naive aggregation functions in GCNs.
• To propose a novel aggregation strategy that enhances node information retention and feature discrimination.
• To improve the theoretical expressivity and practical performance of GCNs.

Main Methods:

• Introduced Adaptive Grouping Aggregation (AGA), inspired by the Weisfeiler-Lehman (WL) Test's label histogram.
• Developed a grouping mechanism using a modified Student's t-Distribution between node features and learnable group labels.
• Implemented the AGA strategy as an end-to-end trainable pipeline using Gumbel Softmax for seamless integration into GCN architectures.

Main Results:

• The AGA strategy significantly enhances feature discrimination by retaining more comprehensive node information.
• Experiments on multiple benchmarks demonstrate consistent performance improvements across all control groups compared to other aggregation strategies.
• AGA achieved state-of-the-art results in most experimental settings, including large-scale benchmarks.

Conclusions:

• The proposed Adaptive Grouping Aggregation (AGA) effectively overcomes the limitations of naive aggregation functions in GCNs.
• AGA offers a powerful and flexible plug-in module that demonstrably improves GCN performance and expressivity.
• The method's superiority is validated through extensive experiments and comparisons with existing state-of-the-art approaches.