MATH · IN · MODELS

Graph Convolutional Network (Kipf & Welling)

research/foundational-architecture

Structures found in this family (1)

Dimensional collapse

By model (1)

GCN depth ensemble, 2-24 layers (Zhang, Higham, Deidda & Tudisco)
Dimensional collapse

Observations (1)

Numerical and effective rank of node representations collapse toward 1 with depth in real trained GCN/GAT networks, and correlate with test accuracy far better than Dirichlet energy

Papers

Are We Measuring Oversmoothing in Graph Neural Networks Correctly? (2025)