Averaging graph samples boosts community detection error bounds exponentially
New spectral method reduces error exponentially with just 2-3 graph samples
A new paper accepted at ICANN 2026 tackles community detection in stochastic block models (SBM) by exploiting multiple independent graph samples drawn from the same underlying distribution. The authors, Sie Hendrata Dharmawan and Peter Chin from Boston University, extend a simplified spectral algorithm that preserves independence of adjacency matrix entries. Their key theoretical contribution is a multi-sample analogue of the spectral norm bound on the noise matrix, which propagates through Davis-Kahan subspace angle analysis to yield an exponential reduction in the error bound γ as the number of samples m increases.
Specifically, they prove that averaging m samples before spectral partitioning can find a γ-correct partition with probability 1-o(1) whenever the signal-to-noise ratio (a-b)²/(a+b) exceeds (C/m) log(2/γ), improving the single-sample requirement by a factor of m. Experimental validation across graph sizes n up to 1000 and sample counts m up to 9 confirms the bounds are sharp, with dramatic improvements in recovery accuracy even with just two or three samples. This rigorous theoretical foundation directly supports graph data augmentation strategies used in modern graph representation learning, offering a principled explanation for why averaging multiple graph views improves downstream task performance.
- Using m independent graph samples reduces the required signal strength for community detection by a factor of m, with error bound γ decaying exponentially in m.
- Theoretical bounds validated on graphs up to n=1000 nodes and up to m=9 samples, showing even 2-3 samples yield dramatic accuracy improvements.
- Provides rigorous theoretical underpinning for graph data augmentation techniques commonly used in modern graph neural network training.
Why It Matters
Rigorous theory proving that averaging multiple graph samples exponentially improves community detection, directly benefiting graph representation learning pipelines.