14
7

Optimal Network Membership Estimation Under Severe Degree Heterogeneity

Abstract

Real networks often have severe degree heterogeneity. We are interested in studying the effect of degree heterogeneity on estimation of the underlying community structure. We consider the degree-corrected mixed membership model (DCMM) for a symmetric network with nn nodes and KK communities, where each node ii has a degree parameter θi\theta_i and a mixed membership vector πi\pi_i. The level of degree heterogeneity is captured by Fn()F_n(\cdot) -- the empirical distribution associated with nn (scaled) degree parameters. We first show that the optimal rate of convergence for the 1\ell^1-loss of estimating πi\pi_i's depends on an integral with respect to Fn()F_n(\cdot). We call a method optimally adaptive to degree heterogeneity\textit{optimally adaptive to degree heterogeneity} (in short, optimally adaptive) if it attains the optimal rate for arbitrary Fn()F_n(\cdot). Unfortunately, none of the existing methods satisfy this requirement. We propose a new spectral method that is optimally adaptive, the core idea behind which is using a pre-PCA normalization to yield the optimal signal-to-noise ratio simultaneously at all entries of each leading empirical eigenvector. As one technical contribution, we derive a new row-wise large-deviation bound for eigenvectors of the regularized graph Laplacian.

View on arXiv
Comments on this paper