pith. sign in

arxiv: 2405.20936 · v6 · submitted 2024-05-31 · 📊 stat.ME

Bayesian Deep Generative Models for Multiplex Networks with Multiscale Overlapping Clusters

Yuren Zhou , Yuqi Gu , David B. Dunson This is my paper

Pith reviewed 2026-05-24 00:45 UTC · model grok-4.3

classification 📊 stat.ME
keywords Bayesian hierarchical modelmultiplex networksoverlapping clustersmultiscale clusteringmodel identifiabilityposterior consistencybrain connectome
0
0 comments X

The pith

A Bayesian hierarchical model infers multiscale overlapping clusters in multiplex networks and proves identifiability plus posterior consistency.

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

The paper develops a generative model for multiple network samples observed on the same nodes, such as brain scans or trade relations. It places a hierarchical structure on the nodes at the population level while allowing each individual network to have its own multiresolution clustering. The authors supply new tools to establish that the model parameters are identifiable, prove that the posterior concentrates on the true values as data grow, and give practical algorithms for drawing posterior samples. These steps let the model separate shared population patterns from sample-specific groupings without hand-tuned resolutions.

Core claim

The authors introduce a Bayesian hierarchical generative process whose latent variables encode multiscale overlapping clusters both across the population of networks and within each replicate; they prove identifiability of all parameters via novel technical arguments, establish posterior consistency, and supply efficient posterior computation procedures that recover the population hierarchy and the replicate-specific cluster assignments.

What carries the argument

Bayesian hierarchical generative model with multiscale overlapping clusters defined at both population and replicate levels.

If this is right

  • Population-level node hierarchy and replicate-level multi-resolution clusters can be inferred jointly from the same data.
  • The identifiability tools apply directly to other hierarchical network models that share similar latent cluster structures.
  • Posterior samples yield uncertainty quantification for both the shared hierarchy and the per-network cluster assignments.
  • The computation methods scale to real multiplex datasets such as brain connectomes.

Where Pith is reading between the lines

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

  • The same identifiability arguments could be reused to prove consistency for models that add node covariates or edge weights.
  • If the model is correct, the inferred population hierarchy supplies a natural way to align clusters across different studies of the same node set.
  • The multiscale structure suggests testing whether coarser or finer resolutions dominate prediction error on held-out networks.

Load-bearing premise

The observed networks are produced exactly by the hierarchical generative process that places multiscale overlapping clusters at both the population and the replicate levels.

What would settle it

Generate synthetic multiplex data from the model with known population hierarchy and replicate clusterings, then check whether the posterior sampler recovers those exact structures with high probability as the number of replicates increases.

Figures

Figures reproduced from arXiv: 2405.20936 by David B. Dunson, Yuqi Gu, Yuren Zhou.

Figure 1
Figure 1. Figure 1: A simple illustration of our model. (a) a graphical model with a three-layer population [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of node adjacency and community adjacency in three individuals, repre [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An illustration of our model, where the Bayesian network [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Posterior distributions of the higher-level and lower-level clusters given by connection [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The distributions of some cognitive traits in the two clusters of individuals with entrywise [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: An illustration of the additive effects of Γ [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: From top to bottom are the trace plot of the Hamming distance [PITH_FULL_IMAGE:figures/full_fig_p097_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: True values of connection matrices A1, A2. We have conducted model selection using WAIC as the information criterion. Fixing K = 2, for each valid choice of p0 and p1, we start with spectral initialization, run 10,000 iterations of subsampling Gibbs sampler with 1% subset ratio, and follow by 100 iterations of standard Gibbs sampler. We compute the WAIC for each choice of p0, p1 and visualize in [PITH_FUL… view at source ↗
Figure 9
Figure 9. Figure 9: WAIC for each choice of network structure with [PITH_FULL_IMAGE:figures/full_fig_p099_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: True value, spectral initialization, and posterior means of the connection matrices [PITH_FULL_IMAGE:figures/full_fig_p100_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: True values and posterior means of the continuous parameters [PITH_FULL_IMAGE:figures/full_fig_p101_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ratio of W1 approximation error over true value for each entry of Γ1,Γ2. perturbations of moderate size, i.e. ±1 shifts for location parameters and ×2 or × 1 2 scalings for variance and concentration parameters. Under each perturbed configuration, we refit the model using the same data via the standard Gibbs sampler and assess the influence on posterior inference using several summary metrics: the posteri… view at source ↗
Figure 14
Figure 14. Figure 14: Recovery accuracy of latent adjacency matrices [PITH_FULL_IMAGE:figures/full_fig_p103_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: WAIC over the p0 − p1 plane for sparsity hyperparameter S ∈ {1, 2}. We visualize the computed WAIC for each choice of p0, p1 under sparsity hyperparameter S = 1, 2 in [PITH_FULL_IMAGE:figures/full_fig_p107_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Posterior distribution of the higher-level and lower-level clusters given by connection [PITH_FULL_IMAGE:figures/full_fig_p108_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Histograms of p-values in Hotelling’s two-sample T 2 tests. A significant p-value represents an active entry (i, j) in X1 that generates an interpretable clustering of the population. I.4 Relating Inferred Latent Features to Cognitive Traits We next relate individual’s cognitive traits to their brain connectivity. We use 33 categories of cognitive traits, including language skills measured by oral reading… view at source ↗
Figure 18
Figure 18. Figure 18: The out-of-sample R2 ’s are presented for ridge regressions that predict different (Gaussian-transformed) dominant principal components of cognitive traits. Four groups of pre￾dictors are considered, which are the latent features in X (n) 1 of our model with S = 1 and S = 2, the latent features in Z (n) of the DCMM with q = 18 and q = 21. We evaluate how the latent brain connectivity features learned from… view at source ↗
Figure 19
Figure 19. Figure 19: ROC curves for predicting masked edges in the brain connectivity networks of 1065 [PITH_FULL_IMAGE:figures/full_fig_p112_19.png] view at source ↗
read the original abstract

Our interest is in multiplex network data with multiple network samples observed across the same set of nodes. Examples originate from a variety of fields, including brain connectivity, international trade networks, and social networks, among others. Our goal is to infer a hierarchical structure of the nodes at a population level, while performing multi-resolution clustering of the individual replicates. To accomplish this, we propose a Bayesian hierarchical model, provide theoretical support in terms of identifiability and posterior consistency, and design efficient methods for posterior computation. We provide novel technical tools for proving model identifiability, which are of independent interest. Our proposed methodology is demonstrated through numerical simulation and an application to brain connectome data.

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

2 major / 1 minor

Summary. The manuscript proposes a Bayesian hierarchical model for multiplex network data observed across multiple samples on the same nodes. The model aims to recover a hierarchical population-level structure on the nodes together with multi-resolution overlapping cluster assignments for the individual replicates. It asserts theoretical support via novel tools establishing model identifiability and posterior consistency, supplies efficient posterior computation methods, and illustrates the approach on numerical simulations and brain connectome data.

Significance. If the identifiability and consistency results can be rigorously established, the work would supply a principled Bayesian framework for multiscale analysis of multiplex networks, addressing overlapping clusters at both population and replicate levels. The claimed novel technical tools for identifiability proofs could be of independent methodological interest beyond the network setting. The brain-connectome application, if accompanied by quantitative validation, would demonstrate practical utility in neuroscience.

major comments (2)
  1. [Abstract] Abstract: the claims of identifiability, posterior consistency, and efficient computation are asserted without any derivation details, proof sketches, or data-exclusion rules; because these are the central theoretical contributions, the manuscript must supply explicit statements of the assumptions, key steps, and any novel technical tools in a dedicated theory section.
  2. [Application] Application section: the brain connectome analysis is mentioned without any quantitative results, error bars, or comparison metrics, so it is impossible to assess whether the model recovers meaningful multiscale structure on real data; this demonstration is load-bearing for the practical claim.
minor comments (1)
  1. [Abstract] The abstract should explicitly reference the section numbers containing the identifiability proofs and the posterior-consistency theorem.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and commit to revisions that strengthen the presentation of the theoretical results and the real-data application.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claims of identifiability, posterior consistency, and efficient computation are asserted without any derivation details, proof sketches, or data-exclusion rules; because these are the central theoretical contributions, the manuscript must supply explicit statements of the assumptions, key steps, and any novel technical tools in a dedicated theory section.

    Authors: The manuscript already contains a theory section (Section 3) that states the model assumptions, outlines the novel technical tools for identifiability, and provides the key steps and proof sketches for both identifiability and posterior consistency. The abstract is kept concise per standard journal length limits. We will revise the abstract to include a brief, explicit statement of the main assumptions, the novel tools, and the consistency result. We will also add a short subsection in the theory section that consolidates the proof strategy and data-exclusion rules for the consistency theorem. revision: yes

  2. Referee: [Application] Application section: the brain connectome analysis is mentioned without any quantitative results, error bars, or comparison metrics, so it is impossible to assess whether the model recovers meaningful multiscale structure on real data; this demonstration is load-bearing for the practical claim.

    Authors: We agree that the current application section provides only a qualitative illustration. In the revised manuscript we will add quantitative validation: posterior summaries of the recovered hierarchical structure and multi-resolution clusters with credible intervals, numerical comparison metrics against baseline multiplex clustering methods, and an assessment of alignment with known neuroanatomical partitions. These additions will allow readers to evaluate the practical utility of the model on the connectome data. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper proposes a Bayesian hierarchical model for multiplex networks, claiming identifiability and posterior consistency via novel technical tools presented as independent contributions. No load-bearing step is shown to reduce by the paper's own equations to a fitted quantity, self-citation chain, or definitional equivalence; the generative assumption is the standard one required for any such consistency result and does not create internal circularity. The derivation chain is self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Based solely on abstract; model rests on standard Bayesian hierarchical assumptions and a generative process for multiplex networks with multiscale clusters. No free parameters or invented entities are quantifiable from the abstract alone.

axioms (1)
  • domain assumption Standard Bayesian hierarchical modeling assumptions including prior distributions on cluster assignments and network edges
    Invoked implicitly to define the generative model and enable posterior inference.
invented entities (1)
  • Multiscale overlapping cluster structure no independent evidence
    purpose: To represent hierarchical population structure and replicate-specific multi-resolution clusters in the generative process
    Core modeling device introduced to capture the multiplex network features described.

pith-pipeline@v0.9.0 · 5643 in / 1248 out tokens · 29246 ms · 2026-05-24T00:45:09.958161+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Local Independence: High-Dimensional Latent Class Graphical Models with Shared Block Structure

    stat.ME 2026-06 unverdicted novelty 6.0

    Introduces latent class graphical models with shared block structure for high-dimensional ordinal responses, a three-step estimator, and finite-sample consistency guarantees under high-dimensional scaling.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    Using the formula (⊗c i=1Ci) (⊗c i=1Di) (⊙c i=1Ei) =⊙ c i=1(CiDiEi) for matricesC i,D i,E i of compat- ible dimensions, we obtain Λ′

    :=⊙ (i,j)∈I1Λ′ (i,j). Using the formula (⊗c i=1Ci) (⊗c i=1Di) (⊙c i=1Ei) =⊙ c i=1(CiDiEi) for matricesC i,D i,E i of compat- ible dimensions, we obtain Λ′

    work page

  2. [2]

    = 1−λ ⋆ 0 1 ⊗ pk−1 (pk−1 −1) 2 1 0 1 1 ⊗ pk−1 (pk−1 −1) 2 Λ[1],(B.18) whereD ⊗c of a matrixDdenotes the Kronecker product⊗ i∈[c]D. On the right hand side of (B.18), we note that the first term is a upper-triangular square matrix with diagonal all ones and the second term is a lower-triangular square matrix with diagonal all ones, which are both invertible...

    work page

  3. [3]

    is invertible. 36 We define thelexicographical orderbetween any two distinct adjacency matricesX k−1 and eXk−1 asX k−1 ≻lex eXk−1 if and only if there exists (s, t)∈ ⟨p k−1⟩such that Xk−1,s,t > eXk−1,s,t, X k−1,i,j = eXk−1,i,j ∀(i, j)∈ (i, j)∈ ⟨p k−1⟩:i < sor (i=s, j < t) . (B.19) If we vectorize the upper triangular off-diagonal entries (i, j) ofX k−1 in...

    work page

  4. [4]

    is given by (NΛ′ [1])s,t = Y (i,j)∈⟨pk−1⟩ (λ(i,j),1,t −λ ⋆)1X−1 k−1(s)i,j=1 + 1X−1 k−1(s)i,j=0 ,(B.20) whereλ (i,j),1,t denotes the (1, t)th entry ofΛ (i,j). SinceΘ∈ T 0 andA k takes the blockwise form (B.13) withP k =I pk, we have each Γ k,i,j >0 anda k,i =e i fori∈[p k−1], which implies λ(i,j),1,t = exp Ck +a ⊤ k,i(Γk ∗X −1 k−1(t))ak,j 1 + exp Ck +a ⊤ k...

    work page

  5. [5]

    are therefore positive. For 1≤s < t≤ |X k−1|, sinceX −1 k−1(s)≻ lex X−1 k−1(t), there exists (i, j)∈ ⟨p k−1⟩such thatX −1 k−1(s)i,j = 1 andX −1 k−1(t)i,j = 0, which implies (λ(i,j),1,t −λ ⋆)1X−1 k−1(s)i,j=1 + 1X−1 k−1(s)i,j=0 = exp Ck + Γk,i,j X−1 k−1(t)i,j 1 + exp Ck + Γk,i,j X−1 k−1(t)i,j −λ ⋆ = 0. From (B.20), all the upper-triangular off-diagonal entr...

    work page

  6. [6]

    This proves thatNΛ ′

    is lower-triangular with positive diagonal entries. This proves thatNΛ ′

    work page

  7. [7]

    With the same procedure, we can also prove that rank K(Λ[2]) =|X k−1|

    is invertible, as isΛ [1], and rank K(Λ[1]) =|X k−1|. With the same procedure, we can also prove that rank K(Λ[2]) =|X k−1|. We next show that (Λ[3] ⊙Λ [4])diag(v) has Kruskal rank≥2. Since each column ofΛ [3] ⊙Λ [4] sums to one and each entry ofvis positive byΘ∈ T 0 and our model formulation (2.5), it suffices to show thatΛ [3] ⊙Λ [4] does not have two i...

    work page 2009

  8. [8]

    = exp X1,i,j C′ 1 +a ⊤ 1,i(P0(X0)∗Γ ′ 1)a1,j 1 + exp C′ 1 +a ⊤ 1,i(P0(X0)∗Γ ′ 1)a1,j = exp X1,i,j C1 +a ⊤ 1,i(X0 ∗Γ 1)a1,j 1 + exp C1 +a ⊤ 1,i(X0 ∗Γ 1)a1,j =P(X 1,i,j|X0,A 1,Θ 1). SinceP(P 0(X0)|ν′) =P(X 0|ν) is implied by the definition ofν ′, taking summations we have P(X1|A,Θ ′) = X X0∈X0  P(P(X0)|ν′) Y (i,j)∈⟨p1⟩ P(X1,i,j|P(X0),A 1,Θ ′ 1)   = X X0...

    work page 2015

  9. [9]

    See Section B.1 for the definition of Kruskal rank

    This also implies rank(M)≥3, since the rank is never small than the Kruskal rank. See Section B.1 for the definition of Kruskal rank. Proof.The cased= 3 follows since all matrices inM 3 are invertible. We prove ford≥4 by contradiction. Suppose there exists distinct indicesi, j, k∈[d] such that the columnsm i,m j,m k ofM∈ M d are linearly dependent, i.e. a...

    work page 1965

  10. [10]

    ,ξ epofY, we divideξ 2,

    In the SCORE step, for the topepeigenvectorsξ 1, . . . ,ξ epofY, we divideξ 2, . . . ,ξ epentrywise overξ 1 and stack the resultedep−1 vectors by columns into a matrixR∈R p×ep−1. For intuition purposes, we also consider the matrix eRobtained by applying this SCORE step to E[Y|D,Π,Z] instead ofY, since the difference betweenRand eRis proven to be small. It...

    work page

  11. [11]

    This allows us to find out the pure nodes of the model

    In the vertex hunting step, we identify the vertices of this simplex using existing methods such as successive projection (Ara´ ujo et al., 2001; Nascimento and Dias, 2005). This allows us to find out the pure nodes of the model. To improve the robustness of this step, it is often helpful to conductk-means clustering of the rows inRbefore finding the vert...

    work page 2001

  12. [12]

    The mixed membershipsΠcan be exactly recovered when the Mixed-SCORE algorithm is applied toE[Y|D,Π,Z]

    In the membership reconstruction step, given the pure nodes, the mixed memberships of each node can be recovered through solving linear equation systems. The mixed membershipsΠcan be exactly recovered when the Mixed-SCORE algorithm is applied toE[Y|D,Π,Z]. In practice, onlyYis known and the Mixed-SCORE algorithm is applied toY, but the estimator ofΠobtain...

    work page 2015

  13. [13]

    Detailed results reported in Section H.2 provide evidence that the Markov chains of Gibbs samplers under each choice ofp k have converged

    and the Geweke statistic (Geweke, 1992) for convergence diagnosis of both the subsampling and the standard Gibbs samplers. Detailed results reported in Section H.2 provide evidence that the Markov chains of Gibbs samplers under each choice ofp k have converged. Given a Bayesian network structure specified byp k, at thetth iteration of the standard Gibbs s...

    work page 1992

  14. [14]

    LetC (2) 1 ,

    This represents a fully-connected ternary tree with its root node removed: the 27 nodes at the finest level are partitioned into 9 leaf communities of size 3, and these leaf communities are further grouped into 3 coarser communities of size 9. LetC (2) 1 , . . . ,C(2) 9 denote the 9 leaf communities of size 3, and letC (1) 1 , . . . ,C(1) 3 denote the coa...

    work page 2022

  15. [15]

    between the latent features and the grouped cognitive measures, with results summarized in Table 3. Second, to assess the practical predictive value of the latent representations, we fit ridge regression models that use the latent features to forecast individual cognitive measures, tuning the regularization parameter by 10-fold cross validation and evalua...

    work page 2023