Gauge-Equivariant Graph Networks via Self-Interference Cancellation

Overview

We propose Gauge-Equivariant Graph Networks via Self-Interference Cancellation (GESC), a novel $\mathrm{U}(1)$-valued transport followed by a rank-1 projection that explicitly removes self-parallel components before attention.

Execution

To train GESC on benchmark datasets (e.g., Cora, Citeseer, Pubmed, Chameleon, Squirrel, Actor, Cornell, Texas, Wisconsin) from 0 to 8:

• python gesc.py 0 # Cora
• python gesc.py 1 # Citeseer
• python gesc.py 2 # Pubmed
• ...
• python gesc.py 8 # Wisconsin