Shift Aggregate Extract Networks.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Network Analysis

Background:

• Learning effective representations for large graphs is challenging due to high variability.
• Existing methods like recursive neural networks struggle with complex graph structures.

Purpose of the Study:

• To introduce a new architecture for learning graph representations using deep hierarchical decompositions.
• To enable the modeling of nested part-of-part relations in graphs.
• To address the limitations of current graph classification methods on large, variable datasets.

Main Methods:

• Extending classic R-decompositions to create a deep hierarchical structure.
• Unrolling a neural network template over the decomposition hierarchy.
• Utilizing domain compression to reduce computational complexity by exploiting symmetries.

Main Results:

• The proposed architecture outperforms state-of-the-art graph classification methods on large social network datasets.
• The method demonstrates competitive performance on small chemobiological benchmark datasets.
• The approach effectively handles high degree variability in social network graphs.

Conclusions:

• Deep hierarchical decompositions offer a powerful framework for graph representation learning.
• This architecture provides a scalable and effective solution for graph classification tasks.
• The method advances the state-of-the-art in both social network analysis and chemobiological data processing.