Embedding Birth-Death Processes within a Dynamic Stochastic Block Model

Statistical clustering in dynamic networks aims to identify groups of nodes with similar or distinct internal connectivity patterns as the network evolves over time. While early research primarily focused on static Stochastic Block Models (SBMs), recent advancements have extended these models to handle dynamic and weighted networks, allowing for a more accurate representation of temporal variations in structure. Additional developments have introduced methods for detecting structural changes, such as shifts in community membership. However, limited attention has been paid to dynamic networks with variable population sizes, where nodes may enter or exit the network. To address this gap, we propose an extension of dynamic SBMs (dSBMs) that incorporates a birth-death process, enabling the statistical clustering of nodes in dynamic networks with evolving population sizes. This work makes three main contributions: (1) the introduction of a novel model for dSBMs with birth-death processes, (2) a framework for parameter inference and prediction of latent communities in this model, and (3) the development of an adapted Variational Expectation-Maximization (VEM) algorithm for efficient inference within this extended framework.

💡 Research Summary

The paper addresses a notable gap in the literature on dynamic network modeling: most existing stochastic block models (SBMs) assume a fixed set of vertices, even when the edges evolve over time. Real‑world systems—such as ecological populations, academic collaborations, or corporate teams—often experience births, deaths, hires, and departures, causing the number of nodes to change. To capture this phenomenon, the authors introduce the Birth‑Death Stochastic Block Model (BD‑SBM), a continuous‑time extension of dynamic SBMs that explicitly couples a homogeneous birth‑death process with a block‑structured edge generation mechanism.

Key model features:

1. Population dynamics – Each node is born according to a Poisson birth process with rate λ and dies with rate μ, both shared across communities. When a birth occurs, the new node inherits the community label of its parent; labels are immutable for the node’s entire lifespan. This reflects settings where lineage or team affiliation is stable (e.g., a PhD student remains in the advisor’s research group until exit).
2. Edge generation – At discrete observation times (snapshots), the current set of alive nodes forms an undirected graph whose edges are independent Bernoulli variables. The success probability for an edge between a node in community k and a node in community m is π_{km}. Thus, each snapshot follows a classic SBM conditional on the latent community assignments that are fixed throughout the observation window.
3. Observed data – The model assumes that (i) the full sequence of birth and death event times and their nature (birth or death) is observed, (ii) the adjacency matrices at a finite set of snapshot times are observed, and (iii) the community labels are latent.

Direct maximum‑likelihood estimation is infeasible because the latent label space grows exponentially with the total number of individuals ever present. The authors therefore adopt a mean‑field variational inference framework. They define a structured variational family that factorises over individual labels Z_i and over community‑size indicators L_{ℓk} at each birth‑death jump time ℓ. The variational distribution for Z_i is a categorical vector, while L_{ℓk} is modeled as a Bernoulli (or multinomial) variable encoding whether community k has a given size at time ℓ.

The inference proceeds via a Variational Expectation‑Maximisation (VEM) algorithm:

• E‑step – Given current parameters θ = (λ, μ, π, β), update the variational factors. The update for each Z_i mirrors the standard SBM variational E‑step, using expected edge counts weighted by the current community‑size variational parameters. The update for L_{ℓk} exploits the fact that a birth or death changes a community size by at most one, yielding simple closed‑form expressions for the posterior probability of each size transition.

• M‑step – Maximise the expected complete‑data log‑likelihood with respect to θ. Because λ and μ appear only in the likelihood of the observed birth‑death events, their M‑step updates reduce to λ̂ = (total births) / (total exposure time) and μ̂ = (total deaths) / (total exposure time), i.e., simple rate estimators. The block‑connectivity matrix π and the initial community proportion β are updated exactly as in static SBM variational EM, using the expected counts of within‑ and between‑community edges.

The algorithm enjoys a monotonic increase of the variational lower bound and converges in a modest number of iterations (typically <20). Computational complexity scales linearly with the number of observed nodes, snapshots, and communities (O(N T K)), making it practical for medium‑size dynamic networks.

Empirical evaluation comprises two parts. First, synthetic experiments vary λ/μ ratios, community separation, and initial population size. The BD‑SBM accurately recovers both latent community assignments (high Normalised Mutual Information) and the underlying birth‑death rates (low relative error), outperforming baseline dynamic SBMs that ignore population changes. Second, the model is applied to a real‑world arXiv collaboration network. Nodes represent authors; edges indicate co‑authorship within yearly snapshots. Observed birth‑death events correspond to authors entering (first publication) or exiting (no further publications). The BD‑SBM identifies coherent research groups (e.g., probability, statistics, machine learning) and captures their growth and shrinkage over years, with estimated λ and μ reflecting known trends in academic hiring and retirement.

Limitations and extensions are discussed. The current formulation assumes homogeneous birth‑death rates across communities and immutable community membership, which precludes modeling scenarios where individuals switch teams or where certain groups have higher turnover. The authors outline possible extensions: (i) community‑specific λ_k, μ_k requiring additional variational parameters, (ii) a Markovian label transition process to allow limited switching, and (iii) a fully Bayesian treatment with hierarchical priors to regularise the increased parameter space.

In summary, the paper makes three substantive contributions: (1) a novel dynamic SBM that integrates a continuous‑time birth‑death process, (2) a structured variational EM algorithm that jointly infers population dynamics, edge‑generation parameters, and latent community structure, and (3) empirical validation showing that the model captures both structural and demographic evolution in synthetic and real networks. By bridging the gap between network topology and population turnover, the BD‑SBM provides a versatile statistical tool for a wide range of disciplines where node birth and death are intrinsic to the system’s dynamics.