C3E connects graph structure, propagation, hidden dimension, and depth through an information-capacity formulation. The result is an auditable architecture-selection mechanism rather than a broad brute-force search.
← All work
C3E — GNN Capacity Estimation
Estimating hidden width and propagation depth from a channel-capacity view of graph representation learning.
Problem
GNN width and depth are often chosen heuristically even though both control information compression and over-squashing.
Approach
Model spectral GNN propagation as a communication channel and estimate capacity-constrained architecture settings through nonlinear optimization.
Role
First author and lead implementation author