Uniform Consistency in Stochastic Block Model with Continuous Community Label

TL;DR

This paper extends the consistency analysis of stochastic block models to cases with continuous community labels, demonstrating uniform consistency of the MLE under certain conditions and exploring discretization effects.

Contribution

It generalizes the existing framework to continuous community labels and analyzes the uniform consistency of MLE in this setting, including the impact of discretization.

Findings

Uniform consistency of MLE in continuous SBM established.

Discretization of community labels achieves similar consistency conditions.

Stronger conditions are required for uniform consistency in continuous case.

Abstract

\cite{bickel2009nonparametric} developed a general framework to establish consistency of community detection in stochastic block model (SBM). In most applications of this framework, the community label is discrete. For example, in \citep{bickel2009nonparametric,zhao2012consistency} the degree corrected SBM is assumed to have a discrete degree parameter. In this paper, we generalize the method of \cite{bickel2009nonparametric} to give consistency analysis of maximum likelihood estimator (MLE) in SBM with continuous community label. We show that there is a standard procedure to transform the $||\cdot||_2$ error bound to the uniform error bound. We demonstrate the application of our general results by proving the uniform consistency (strong consistency) of the MLE in the exponential network model with interaction effect. Unfortunately, in the continuous parameter case, the condition ensuring uniform consistency we obtained is much stronger than that in the discrete parameter case, namely $n\mu_n^5/(\log n)^{8}\rightarrow\infty$ versus $n\mu_n/\log n\rightarrow\infty$. Where $n\mu_n$ represents the average degree of the network. But continuous is the limit of discrete. So it is not surprising as we show that by discretizing the community label space into sufficiently small (but not too small) pieces and applying the MLE on the discretized community label space, uniform consistency holds under almost the same condition as in discrete community label space. Such a phenomenon is surprising since the discretization does not depend on the data or the model. This reminds us of the thresholding method.