Learning Graph Representations With Maximal Cliques.

Area of Science:

• Graph representation learning
• Deep learning on graphs
• Machine learning

Background:

• Non-Euclidean graph structures pose challenges for deep learning.
• Graph convolutional networks (GCNs) are effective for graph classification but have performance limitations.
• Existing GCNs often use traditional neighborhood averaging schemes.

Purpose of the Study:

• Introduce a novel representation learning approach for graph-structured data.
• Improve node representation learning in a semisupervised setting.
• Address limitations of traditional GCN neighborhood aggregation.

Main Methods:

• Developed COnvOlving cLiques (COOL), a neighborhood aggregation method.
• Utilized spectral convolutions on graph-structured data.
• Employed maximal clique identification for local information aggregation.

Main Results:

• COOL aggregates information from densely connected neighbors, considering differing locality.
• The method significantly improves performance on multiple transductive node classification tasks.
• Achieved substantial gains over existing graph neural network approaches.

Conclusions:

• COOL offers a more effective neighborhood aggregation strategy for GCNs.
• The approach enhances node representation learning for graph data.
• Demonstrated superior performance in semisupervised transductive node classification.