Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications

In this paper we extend our previous work on the stochastic block model, a commonly used generative model for social and biological networks, and the problem of inferring functional groups or communities from the topology of the network. We use the cavity method of statistical physics to obtain an asymptotically exact analysis of the phase diagram. We describe in detail properties of the detectability/undetectability phase transition and the easy/hard phase transition for the community detection problem. Our analysis translates naturally into a belief propagation algorithm for inferring the group memberships of the nodes in an optimal way, i.e., that maximizes the overlap with the underlying group memberships, and learning the underlying parameters of the block model. Finally, we apply the algorithm to two examples of real-world networks and discuss its performance.

💡 Research Summary

This paper presents a comprehensive asymptotic analysis of the stochastic block model (SBM) and derives an optimal algorithm for community detection and parameter learning. Building on the authors’ earlier work, the study uses the cavity method from statistical physics to compute the free‑energy density of the SBM in the thermodynamic limit (N → ∞) for sparse graphs where the average degree remains O(1). The SBM is defined by the number of groups q, the group size fractions {n_a}, and the affinity matrix {p_ab}. By scaling p_ab = c_ab / N, the model captures realistic sparse networks while keeping the average degree c = Σ_{ab} c_ab n_a n_b finite.

The authors first formulate the Bayesian inference problem: given a graph G generated by an SBM with unknown parameters θ = {q, {n_a}, {p_ab}}, (i) estimate the most likely parameters (parameter learning) and (ii) infer the most likely group assignment for each node (inference). They introduce a normalized overlap Q that measures correlation between the inferred labeling and the planted ground truth; Q = 1 indicates perfect recovery, while Q = 0 corresponds to random guessing.

Using the cavity method, the posterior distribution over labelings is shown to be equivalent to the Boltzmann distribution of a generalized Potts model with Hamiltonian H({q_i}|G,θ) = – Σ_i log n_{q_i} – Σ_{i<j}