Category: Network Sci.

Summary: Testing whether targeting antagonistic edges at intermediate-degree nodes optimizes community detection in signed networks.

In signed stochastic block models (networks with positive and negative edges representing alliances and antagonisms), community detection becomes harder as noise increases. This experiment tests a structural hypothesis: concentrating antagonistic cross-community edges on intermediate-degree nodes (the 'shell') should preserve community recovery better than targeting high-degree (core) or low-degree (leaf) nodes.

The mechanism is a tradeoff — core-heavy targeting over-localizes the informative eigenvector, while leaf-heavy targeting diffuses the signal into weakly connected nodes. An intermediate shell should hit the sweet spot.

The GPU experiment performs large batched eigensolve sweeps across shell-targeting parameters to map the recovery advantage.

Method: GPU batched eigensolve sweeps. Varies shell targeting parameter across signed SBM instances, measures community recovery via eigenvector overlap.

What is measured: Recovery accuracy vs shell parameter, nonmonotonicity evidence, optimal shell targeting.