[Paper Review] Stochastic Block Model and Community Detection in the Sparse Graphs: A spectral algorithm with optimal rate of recovery

Peter Chin, Anup Rao+1 more|arXiv (Cornell University)|Jan 20, 2015

Complex Network Analysis Techniques25 references94 citations

TL;DR

This paper presents a simple spectral algorithm for community detection in sparse stochastic block models with k blocks, achieving optimal recovery rates. It proves that under the condition $(a-b)^2/(a+b) \geq C \log(1/\gamma)$, the algorithm recovers $\gamma$-correct partitions with high probability, settling an open question on optimal recovery thresholds for sparse graphs.

ABSTRACT

In this paper, we present and analyze a simple and robust spectral algorithm for the stochastic block model with $k$ blocks, for any $k$ fixed. Our algorithm works with graphs having constant edge density, under an optimal condition on the gap between the density inside a block and the density between the blocks. As a co-product, we settle an open question posed by Abbe et. al. concerning censor block models.

Motivation & Objective

• To develop a simple, robust spectral algorithm for community detection in sparse stochastic block models with $k$ blocks.
• To establish the optimal recovery threshold for community detection in sparse graphs, where edge probabilities are constants.
• To resolve an open question on the minimal condition required for $\gamma$-correct recovery in the stochastic block model.
• To extend the analysis to general $k$-block models with possibly unequal block sizes.
• To prove that the derived condition is tight up to constant factors, matching minimax lower bounds.

Proposed method

• The algorithm uses spectral clustering on the adjacency matrix of the random graph, leveraging the eigenvectors corresponding to the largest eigenvalues to recover community structure.
• It applies a thresholding and rounding procedure on the leading eigenvectors to assign nodes to communities.
• The analysis relies on concentration inequalities and spectral norm bounds for random matrices with block structure.
• The proof partitions the analysis into cases based on node degrees and eigenvalue decay to control error terms.
• It uses a careful bookkeeping of constants and applies geometric series bounds to control sums over node pairs.
• The method is robust to block size imbalances and extends naturally to unequal block sizes with minor modifications.

Experimental results

Research questions

• RQ1 What is the minimal condition on the difference between intra- and inter-block edge probabilities that allows for $\gamma$-correct community recovery in sparse stochastic block models?
• RQ2 Can a simple spectral algorithm achieve the optimal recovery rate without requiring strong assumptions on $a$ and $b$?
• RQ3 Is the $\log(1/\gamma)$ dependence in the recovery threshold tight, or can it be improved?
• RQ4 How does the algorithm perform in the general $k$-block setting with fixed $k$?
• RQ5 Can the algorithm handle unequal block sizes while maintaining optimal recovery guarantees?

Key findings

• For $k=2$, the spectral algorithm achieves $\gamma$-correct recovery with high probability when $(a-b)^2/(a+b) \geq C \log(1/\gamma)$, matching the conjectured optimal threshold.
• The algorithm is simple, robust, and runs in polynomial time, improving upon prior complex algorithms.
• For general $k$ blocks, the condition $(a-b)^2 \geq C_2 k^2 a \log(1/\gamma)$ ensures $\gamma$-correct recovery with high probability.
• The result is tight up to constant factors, as minimax lower bounds suggest that $\Omega(k^2 a)$ scaling is necessary.
• The method extends to unequal block sizes as long as sizes are comparable, with constants depending on the minimum size ratio.
• The paper resolves an open question by Abbe et al. regarding censor block models, confirming the optimality of the $\log(1/\gamma)$ term.