pith. sign in

arxiv: 2604.20266 · v3 · submitted 2026-04-22 · 📊 stat.ME

Bayesian Modeling of the Stochastic Block Model for Weighted Network Data with Zero-Inflated Negative Binomial Distribution

Fumiya Iwashige This is my paper

Pith reviewed 2026-05-10 00:11 UTC · model grok-4.3

classification 📊 stat.ME
keywords stochastic block modelzero-inflated negative binomialweighted networksBayesian inferencecommunity detectioncovariate effectsmissing link predictionoverdispersion
0
0 comments X

The pith

A Bayesian stochastic block model using zero-inflated negative binomial distributions models overdispersed weighted networks with covariates and infers community count from data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces ZINB-SBM and CZINB-SBM as Bayesian models for detecting communities in weighted networks where interaction strengths vary widely and many possible edges are absent. These models replace the restrictive Poisson assumption with a zero-inflated negative binomial that directly captures overdispersion and treats zeros as either structural absences or sampling zeros. Efficient Gibbs sampling is achieved through Polya-Gamma augmentation when pairwise covariates are included, while a dynamic mixture of finite mixtures lets the data determine the number of blocks. Simulations establish greater robustness than zero-inflated Poisson versions when overdispersion is high, and real-data examples show clearer block-specific covariate effects together with higher accuracy in predicting unobserved links. The approach matters for any setting where network weights are counts that exhibit extra variability and incomplete observation, such as social ties, biological interactions, or trade volumes.

Core claim

The authors establish that embedding the zero-inflated negative binomial distribution inside the stochastic block model framework, together with Polya-Gamma data augmentation and a dynamic mixture of finite mixtures, yields posterior samples that recover community structure more reliably under overdispersion, quantify uncertainty in covariate effects per block, and produce superior predictions for missing edges compared with Poisson-regression Bayesian stochastic block models.

What carries the argument

Zero-inflated negative binomial likelihood inside the stochastic block model, with Polya-Gamma augmentation for regression coefficients and dynamic mixture of finite mixtures for unknown block count.

If this is right

  • The model recovers community structure accurately even when edge weights display variance much larger than their mean.
  • Pairwise covariates receive block-specific coefficient estimates with full posterior uncertainty.
  • The number of communities is learned from the data rather than supplied in advance.
  • Missing-link prediction improves because zero inflation is modeled explicitly rather than absorbed into a single Poisson rate.
  • Posterior computation remains feasible through standard Gibbs steps even after covariates are added.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same construction could be adapted to directed or time-varying networks by replacing the symmetric block matrix with an appropriate directed or dynamic version.
  • In domains such as protein-interaction or transportation networks, the improved handling of overdispersion may reduce the number of spurious communities that Poisson models tend to create around high-variance hubs.
  • Because missing-link prediction is a direct byproduct of the zero-inflation component, the framework offers a natural Bayesian route to network completion tasks without separate imputation steps.

Load-bearing premise

The weighted edges are generated independently from a zero-inflated negative binomial distribution once the latent community assignments and any covariates are fixed.

What would settle it

A collection of simulated or real weighted networks with known blocks, high overdispersion, and held-out edges on which the ZINB-SBM recovers the blocks less accurately or predicts the held-out edges worse than a zero-inflated Poisson stochastic block model.

Figures

Figures reproduced from arXiv: 2604.20266 by Fumiya Iwashige.

Figure 1
Figure 1. Figure 1: Posterior means (◦ and △) and 95% credible intervals of the CZINB-SBM regression coefficients for the pairwise covariates Genetic, Taxonomic, and Geographic distance. Solid lines represent the coefficients β11,1, β22,1, and β33,1 for within-community block pairs, while dashed lines represent the coefficients β12,1, β13,1, and β23,1 for between-community block pairs. it may instead fit the data by subdividi… view at source ↗
read the original abstract

Weighted networks encode not only the presence of interactions but also their strength. Existing methods for weighted network community detection often rely on Poisson models, which can be restrictive for overdispersed data and make efficient posterior computation difficult when covariates are incorporated. We propose Bayesian stochastic block models based on the zero-inflated negative binomial distribution: ZINB-SBM without covariates and CZINB-SBM with pairwise covariates. The proposed models accommodate overdispersion, naturally account for missing interactions through zero inflation, and admit efficient Gibbs sampling. In CZINB-SBM, P\'{o}lya-Gamma data augmentation enables posterior inference for regression coefficients with uncertainty quantification. We further employ a dynamic mixture of finite mixtures, which allows the number of communities to be inferred from the data and can lead to more accurate clustering. Simulation studies show that ZINB-SBM is more robust than a zero-inflated Poisson SBM for highly overdispersed networks. Real data analysis demonstrates interpretable block specific covariate effects and substantially improved missing link prediction compared with a Poisson regression-based Bayesian SBM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper proposes Bayesian stochastic block models (ZINB-SBM and its covariate extension CZINB-SBM) for weighted networks based on the zero-inflated negative binomial distribution. The models handle overdispersion and zero-inflation for missing interactions, use Pólya-Gamma data augmentation for regression coefficients in the covariate version, and employ a dynamic mixture of finite mixtures prior to infer the number of communities K. Simulations compare robustness to a zero-inflated Poisson SBM under high dispersion, while real-data examples illustrate interpretable block-specific covariate effects and improved missing-link prediction over a Poisson regression Bayesian SBM.

Significance. If the results hold, the work provides a practical and computationally tractable extension of SBMs to overdispersed weighted networks with covariates, addressing a common limitation of Poisson-based models. Credit is due for the explicit use of established Pólya-Gamma augmentation and dynamic MFM prior, which enable efficient Gibbs sampling and automatic inference of K without ad-hoc model selection.

minor comments (3)
  1. The simulation design (Section 4) could more explicitly state the range of dispersion parameters and zero-inflation probabilities used to generate the highly overdispersed networks, to allow readers to assess how far the robustness claim generalizes beyond the reported settings.
  2. In the real-data analysis (Section 5), the comparison of missing-link prediction performance would benefit from reporting both AUC and precision-recall curves or additional baselines (e.g., a non-zero-inflated NB model) to strengthen the claim of 'substantially improved' performance.
  3. Notation for the negative-binomial dispersion parameter and zero-inflation probability should be introduced once in Section 2 and used consistently; occasional redefinition risks minor confusion for readers.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and constructive review of our manuscript. We are pleased that the referee accurately summarizes the contributions of the ZINB-SBM and CZINB-SBM models, recognizes their significance for handling overdispersion and zero-inflation in weighted networks, and highlights the computational advantages of the Pólya-Gamma augmentation and dynamic mixture of finite mixtures prior. We will prepare a revised version in accordance with the minor revision recommendation.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper explicitly defines the ZINB-SBM and CZINB-SBM as generative models using the zero-inflated negative binomial distribution conditional on block assignments and covariates, then applies standard established tools (Pólya-Gamma augmentation for the regression component and dynamic MFM prior for unknown K) whose validity is independent of the target results. Simulation robustness claims and real-data improvements in link prediction are evaluated against external benchmarks (zero-inflated Poisson SBM and Poisson regression SBM) rather than being forced by construction from fitted parameters or self-citations. No derivation step reduces to renaming inputs, smuggling an ansatz via prior work, or invoking a uniqueness theorem that collapses to the authors' own unverified assumptions.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The claims rest on standard conditional independence assumptions of block models and the choice of zero-inflated negative binomial as the data-generating process, with parameters estimated from data.

free parameters (2)
  • Negative binomial dispersion parameter
    Controls extra variation beyond the mean and is estimated from the data.
  • Zero-inflation probability
    Probability of structural zeros, estimated per block or globally.
axioms (2)
  • domain assumption Edges are conditionally independent given community memberships and parameters.
    Fundamental modeling assumption shared with all stochastic block model variants.
  • standard math Prior distributions on parameters allow proper posterior inference.
    Required for the Bayesian framework and Gibbs sampling to be well-defined.

pith-pipeline@v0.9.0 · 5487 in / 1316 out tokens · 47238 ms · 2026-05-10T00:11:57.599483+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    Argiento, R. and M. D. Iorio (2022). Is infinity that far? A Bayesian nonparametric perspective of finite mixture models . The Annals of Statistics \/ 50\/ (5), pp. 2641--2663

    work page 2022

  2. [2]

    Escobar, M. D. and M. West (1995). Bayesian density estimation and inference using mixtures. Journal of the American Statistical Association \/ 90\/ (430), 577--588

    work page 1995

  3. [3]

    u hwirth-Schnatter, S., G. Malsiner-Walli, and B. Gr \

    Fr \"u hwirth-Schnatter, S., G. Malsiner-Walli, and B. Gr \"u n (2021). Generalized Mixtures of Finite Mixtures and Telescoping Sampling . Bayesian Analysis\/ 16 \/ (4), 1279 -- 1307

    work page 2021

  4. [4]

    Ghosh, S. K., P. Mukhopadhyay, and J.-C. Lu (2006). Bayesian analysis of zero-inflated regression models. Journal of Statistical Planning and Inference\/ 136 \/ (4), 1360--1375

    work page 2006

  5. [5]

    Greve, J., B. Grün, G. Malsiner-Walli, and S. Frühwirth-Schnatter (2022). Spying on the prior of the number of data clusters and the partition distribution in Bayesian cluster analysis . Australian & New Zealand Journal of Statistics\/ 64\/ (2), 205--229

    work page 2022

  6. [6]

    Iwashige, F. and S. Hashimoto (2025). Bayesian mixture modeling using a mixture of finite mixtures with normalized inverse Gaussian weights

    work page 2025

  7. [7]

    Lu, C. (2023). Bayesian Strategies for Complex Statistical Models . PhD thesis , University College Dublin, School of Mathematics and Statistics

    work page 2023

  8. [8]

    Durante, and N

    Lu, C., D. Durante, and N. Friel (2025, 03). Zero-inflated stochastic block modelling of efficiency-security trade-offs in weighted criminal networks. Journal of the Royal Statistical Society Series A: Statistics in Society\/ , qnaf029

    work page 2025

  9. [9]

    Robin, and C

    Mariadassou, M., S. Robin, and C. Vacher (2010). Uncovering latent structure in valued graphs: A variational approach . The Annals of Applied Statistics\/ 4 \/ (2), 715 -- 742

    work page 2010

  10. [10]

    Meilă, M. (2007). Comparing clusterings—an information based distance. Journal of Multivariate Analysis\/ 98\/ (5), 873--895

    work page 2007

  11. [11]

    Miller, J. W. and M. T. Harrison (2018). Mixture models with a prior on the number of components. Journal of the American Statistical Association \/ 113\/ (521), 340--356. PMID: 29983475

    work page 2018

  12. [12]

    Neelon, B. (2019). Bayesian zero-inflated negative binomial regression based on P \'o lya-Gamma mixtures . Bayesian Analysis\/ 14\/ (3), 829

    work page 2019

  13. [13]

    Nowicki, K. and T. A. B. Snijders (2001). Estimation and prediction for stochastic blockstructures. Journal of the American Statistical Association \/ 96\/ (455), 1077--1087

    work page 2001

  14. [14]

    Polson, N. G., J. G. Scott, and J. Windle (2013). Bayesian inference for logistic models using P \'o lya--Gamma latent variables . Journal of the American statistical Association \/ 108 \/ (504), 1339--1349

    work page 2013

  15. [15]

    Piou, and M.-L

    Vacher, C., D. Piou, and M.-L. Desprez-Loustau (2008, 03). Architecture of an antagonistic tree/fungus network: The asymmetric influence of past evolutionary history. PLOS ONE\/ 3\/ (3), 1--10

    work page 2008

  16. [16]

    Wade, S. and Z. Ghahramani (2018). Bayesian Cluster Analysis: Point Estimation and Credible Balls (with Discussion) . Bayesian Analysis\/ 13 \/ (2), 559 -- 626

    work page 2018