A spectral method for community detection in moderately-sparse degree-corrected stochastic block models
Résumé
We consider community detection in Degree-Corrected Stochastic Block Models. We perform spectral clustering on $\widehat{H} = \left(\frac{1}{\widehat{D}_i \widehat{D}_j} A_{ij} \right)_{i,j=1}^n,$ where $A$ is the adjacency matrix of the network containing $n$ vertices and $\widehat{D}_i$ is the observed degree of node $i$. We show that this leads to consistent recovery of the block-membership of all but a vanishing fraction of nodes, even when the lowest degree is of order log$(n)$. There turns out to be a natural connection between $\widehat{H}$ and random walks on instances of the random graph. Moreover, $\widehat{H}$ appears to have a better behaved eigenspace than the ordinary adjacency matrix in case of a very heterogeneous degree-sequence.