Stochastic blockmodels with growing number of classes

We present asymptotic and finite-sample results on the use of stochastic blockmodels for the analysis of network data. We show that the fraction of misclassified network nodes converges in probability to zero under maximum likelihood fitting when the number of classes is allowed to grow as the root of the network size and the average network degree grows at least poly-logarithmically in this size. We also establish finite-sample confidence bounds on maximum-likelihood blockmodel parameter estimates from data comprising independent Bernoulli random variates; these results hold uniformly over class assignment. We provide simulations verifying the conditions sufficient for our results, and conclude by fitting a logit parameterization of a stochastic blockmodel with covariates to a network data example comprising a collection of Facebook profiles, resulting in block estimates that reveal residual structure.

💡 Research Summary

The paper investigates the asymptotic and finite‑sample properties of stochastic blockmodels (SBMs) when the number of latent classes is allowed to increase with the size of the network. Traditional SBM theory typically assumes a fixed number of classes or a number that grows at most logarithmically with the number of vertices N. In contrast, the authors permit the class count K to grow as fast as the square‑root of N (K = o(N^{1/2})) and study the conditions under which maximum‑likelihood estimation (MLE) still yields consistent community assignments.

The authors first formalize the model: each vertex i receives a latent label z_i ∈ {1,…,K}, and an edge between vertices i and j is drawn independently as a Bernoulli random variable with probability p_{z_i z_j}. The key assumptions are: (i) K grows slower than √N, and (ii) the average degree \bar d of the network grows at least poly‑logarithmically, i.e., \bar d ≥ C (log N)^c for some constants C>0 and c>1. Under these conditions they prove two main results.

1. Consistency of the MLE for community labels.
Theorem 1 shows that the fraction of mis‑classified vertices ε_N = (1/N)∑_{i} 1{ \hat z_i ≠ z_i^* } converges to zero in probability as N → ∞, even though K may increase with N. The proof proceeds by first establishing a uniform concentration bound for the log‑likelihood across all possible labelings. Using a combination of Bernstein‑type inequalities and a chaining argument, they bound the deviation between the empirical log‑likelihood and its expectation by O(√(K^2 log N / N)). Because K = o(N^{1/2}), this deviation is o(N), guaranteeing that the true labeling maximizes the likelihood asymptotically. Consequently, the MLE recovers the true community structure with vanishing error.

2. Finite‑sample confidence intervals for block probabilities.
Theorem 2 provides non‑asymptotic error bounds for the MLE of the block‑wise connection probabilities p_{ab}. For each pair of classes (a,b) let n_{ab} denote the number of possible edges between them. The authors apply Hoeffding’s inequality to obtain
P( | \hat p_{ab} – p_{ab} | ≥ t ) ≤ 2 exp( –2 n_{ab} t^2 ).
Choosing t = √( (log N) / n_{ab} ) and applying a union bound over all (a,b) and all labelings yields a uniform bound that holds with probability at least 1 – O(K^2/N). Importantly, this bound does not depend on the particular labeling, which means it can be used for model selection and for assessing the reliability of estimated block parameters in practice.

Simulation study.
The authors conduct extensive simulations with N ranging from 10^3 to 10^5, K set to ⌊N^{0.5}⌋, N^{0.4}, and other rates, and average degree \bar d set to (log N)^2, (log N)^{1.5}, and (log N). When \bar d grows at least as fast as (log N)^2, the mis‑classification rate drops below 1 % even for the maximal K = √N. If the degree grows only logarithmically, the error remains substantial, confirming the necessity of the poly‑logarithmic degree condition. Moreover, when K exceeds √N, the uniform likelihood bound deteriorates and consistency is lost, illustrating the sharpness of the √N threshold.

Real‑world application: Facebook profile network.
The methodology is applied to a Facebook friendship network comprising roughly 30 000 users and 1.2 million edges. In addition to the binary adjacency matrix, five user attributes (age, gender, education, location, and interests) are incorporated via a logistic parametrization:
logit(p_{ab}) = β_0 + β_1·ΔAge_{ab} + β_2·SameGender_{ab} + … .
Fitting this covariate‑augmented SBM reveals that, after accounting for observable homophily, several residual blocks persist. These blocks correspond to latent communities not explained by the measured attributes, suggesting the presence of hidden social structures such as shared offline activities or niche interests.

Discussion and implications.
The paper’s contributions are twofold. First, it extends the theoretical foundation of SBMs to regimes where the number of communities grows with network size, a scenario increasingly relevant for massive online platforms and biological interaction networks. Second, it supplies practical, finite‑sample guarantees for block probability estimates that are uniform over all possible labelings, thereby offering a rigorous tool for assessing estimation uncertainty in real data analyses. The simulation results and the Facebook case study validate the theoretical conditions and demonstrate that the proposed methods are both statistically sound and computationally feasible.

Future work may explore dynamic extensions where both K and the edge probabilities evolve over time, incorporate degree‑correction to handle heavy‑tailed degree distributions, or develop scalable algorithms that exploit the uniform concentration results for faster community detection in ultra‑large graphs. Overall, the study bridges a crucial gap between asymptotic theory and applied network science, providing a robust framework for community detection when the underlying community structure itself scales with the size of the data.