Multi-level attention pooling for graph neural networks: Unifying graph representations with multiple localities.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) excel at learning representations from graph-structured data via message passing.
• Deep GNNs struggle with oversmoothing, losing crucial local information while capturing global structures.

Purpose of the Study:

• To propose a novel architecture, Multi-level Attention Pooling (MLAP), for graph-level classification tasks.
• To address the performance degradation in deep GNNs by preserving local and global structural information.

Main Methods:

• Developed MLAP with an attention pooling layer at each message passing step.
• Unified layer-wise graph representations to create a comprehensive graph representation.
• Preserved layer-wise information to mitigate oversmoothing effects.

Main Results:

• MLAP architecture significantly improved graph classification performance over baseline GNNs.
• Experimental results demonstrated enhanced performance compared to traditional architectures.
• Analyses confirmed that multi-level information aggregation improves representation discriminability.

Conclusions:

• MLAP effectively captures graph structural information at multiple locality levels.
• The proposed method enhances GNNs' ability to learn discriminative representations for graph classification.
• MLAP offers a promising approach to overcome limitations of deep GNNs.