{"id":"c91f5611-c072-4726-9834-22521b7110c3","arxiv_id":"2507.09370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LaPCoM is a Bayesian nonparametric mixture-of-mixtures latent position model that jointly clusters networks and nodes in multiplex data.","lead":"This paper introduces LaPCoM, a Bayesian model that simultaneously groups networks in a multiplex dataset and detects communities of nodes within each group of networks. It uses a two-level mixture of latent position models, which could help analysts summarize social, biological, or brain networks by revealing shared structure across networks and roles within them.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LaPCoM's co-clustering claim rests on the assumption that all networks in a network-level cluster share one latent position matrix Z_g; because the simulations generate data from this same assumption, the risk of distorted clusters under within-cluster latent heterogeneity remains untested.","rationale":"The reader's conditional verdict is appropriate, and I agree with the identified weakest assumption. LaPCoM is a coherent and well-specified model, with available code and credible simulation evidence under the model's own generative assumptions. However, the strongest claim—accurate simultaneous network- and node-level clustering—depends on the exchangeability of networks within a cluster, formalized by a single shared Z_g. The paper's simulations never vary this assumption, so they cannot establish robustness to within-cluster latent heterogeneity. The Aarhus application provides some evidence that the assumption is strained: per-network LPCM fits show different node-cluster structure across the four networks assigned to one shared latent space, and the authors make post-hoc choices to obtain the reported interpretable result. The primary school application is more supportive, and the posterior predictive work is a positive feature, but the central methodological risk remains the shared-Z_g assumption. A targeted simulation with network-specific latent perturbations would settle whether this concern is substantive or merely theoretical. Since the reader already judged the paper conditional, my read does not change the verdict.","tokens_in":39764,"tokens_out":8479,"duration_ms":110837,"concrete_test":"Simulate a multiplex in which each network m in a true network-level cluster g has its own latent positions Z_{g,m,i} drawn from the same node-level mixture as the cluster (e.g., z_{g,m,i} ~ N(μ_{g,k(i)}, Σ_{g,k}) for each node i), so within-cluster networks are heterogeneous while sharing cluster-level means and covariances. Fit LaPCoM under the Section 4.2 settings and compare the estimated G+ and network-level ARI to the true G*. If LaPCoM systematically overestimates G+ or drops network-level ARI by more than 0.1 relative to the homogeneous-Z case, the shared-Z_g assumption is the limiting factor for the central claim. Report node-level ARI as well to check whether the compromise latent space distorts the inferred node clusters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the exchangeability assumption introduced in Section 2.2: all networks allocated to a network-level cluster g contribute a likelihood using the same latent positions z_{g,i}, so there is no network-specific latent perturbation, density offset, or sender/receiver effect. If networks in a true cluster have heterogeneous latent positions, the model is forced either to split the cluster (inflating G+) or to estimate a compromise Z_g that blurs node-level clusters. The simulations in Section 4 do not stress this assumption: every scenario generates networks from the fitted model, with a shared Z_g per cluster, so accurate recovery is partly built into the data-generating process. The real-data evidence is suggestive but not decisive: in the Aarhus application (Section 5.2), four networks are forced to share Z_1, while the per-network LPCM fits shown in Figure 7 yield different node-cluster structures (1, 3, 7, and 1 clusters) across those same networks. The post-hoc exclusion of chains and overriding of the merged K=2 solution for the Facebook cluster further weaken this particular demonstration. This does not invalidate the model, but it means the headline claim of simultaneous, accurate two-level clustering is conditional on a shared-latent-space assumption that the paper never varies or tests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LaPCoM, a hierarchical mixture-of-mixtures latent position model for multiplex networks that simultaneously clusters networks and, within each network-level cluster, clusters nodes in a shared latent space. The model uses a mixture of finite mixtures at both levels, with priors on the number of active clusters, and is fitted via a Metropolis-within-Gibbs sampler with telescoping sampling and post-processing for label switching. The authors evaluate the method in two simulation studies, compare it with four existing approaches, and apply it to three real multiplex datasets, reporting posterior predictive checks in each case.","tokens_in":40006,"tokens_out":2825,"duration_ms":37712,"significance":"If the shared-latent-space assumption holds, LaPCoM is a useful contribution: it provides a single interpretable framework for two-level clustering of multiplex networks, supports both binary and count data, and avoids fixing the number of clusters. The supplementary material contains full conditional derivations and MCMC pseudocode, and the R code is publicly available, which are substantial strengths. The simulation study is broad, and the comparison with four competitors gives a useful picture of the model's performance. However, the central exchangeability assumption is not stress-tested, and some real-data conclusions rely on post-hoc decisions, so the strength of the empirical claims is currently stronger than the evidence supports.","major_comments":[{"comment":"The model's central identifying assumption is that all networks in a network-level cluster share a single latent position matrix Z_g, introduced in the likelihood in Section 2.2. The simulation studies in Section 4 generate data from this exact model, so accurate recovery is partly a self-consistency check. The assumption is never varied: there is no scenario in which networks within a true cluster have network-specific latent perturbations, density offsets, or sender/receiver effects. This is load-bearing because under within-cluster latent heterogeneity the model may either split clusters or produce a compromise latent space that blurs node-level clusters. Please add a simulation that generates networks as Z_{g,i} plus a network-specific perturbation, or with network-specific intercept shifts, and report the resulting G+ and node-level ARI.","section":"Section 2.2 and Section 4"},{"comment":"The Aarhus application contains two post-hoc decisions that weaken the illustrative claim. First, two chains with node-level cluster estimates K2+ = 6 and K2+ = 5 are excluded because the K2+ = 5 solution was 'deemed spurious' without a stated quantitative criterion. Second, after post-processing merged the two clusters in the Facebook latent space into a single cluster, the authors override this and 'consider the K2+ = 2 solution more interpretable'. Since the posterior distribution of K2+ reportedly had 'considerable spread', the reported node-level structure should be accompanied by a sensitivity analysis that retains all chains and reports the posterior mass on each K2+, rather than selecting the solution that is most interpretable.","section":"Section 5.2"},{"comment":"The primary-school application shows a clear model deficiency for count-valued multiplexes: the posterior predictive ECDF of positive edge counts systematically underestimates large counts, and the authors state that 'the Poisson distribution may be too restrictive'. Because count-valued networks are one of the two data types the model is claimed to accommodate, this is not a peripheral issue. The manuscript acknowledges this in Section 6, but the main text still presents the analysis as successful. Please either fit a more flexible edge distribution (e.g., zero-inflated Poisson or negative binomial) or substantially temper the claims about count-valued data.","section":"Section 5.3 and Supplementary Section I"},{"comment":"The comparison with PopNet is reported as favorable to LaPCoM, but the tables show PopNet achieving perfect network-level ARI in all five scenarios, while LaPCoM attains 0.86–1.00 with wider credible intervals for G+. The stated advantages of LaPCoM are real (node-level clustering, count data support, lower runtime), but the wording 'matched this accuracy' in the discussion of Table 2 is misleading. Please report the comparison with a more neutral characterization, distinguishing recovery accuracy from model flexibility.","section":"Section 4.2, Tables 2 and 3"}],"minor_comments":[{"comment":"In the introduction, the latent position cluster model is referred to as 'LPM' instead of 'LPCM' when citing Handcock et al. (2007); the abbreviation is inconsistent with the rest of the paper.","section":"Section 1"},{"comment":"There is a typo: 'prescence' should be 'presence'.","section":"Section 5.2"},{"comment":"In the definition of the expected variance V(N), the text states 'nmin = 5 if N < 60, and nmin = 5 if N ≥ 60'; the second condition should presumably be nmin = 10, consistent with Table 1 and the surrounding discussion.","section":"Supplementary Section B"},{"comment":"The axis labels in Figure 9 contain R code fragments such as 'primary_school_pp$primary_school_Z[[optimal_clustering_networks[m]]][,1]'; these should be replaced with clean labels such as 'Dimension 1' and 'Dimension 2'.","section":"Section 5.3 and Figure 9"},{"comment":"The sentence 'the interquartile range (IQR) was 0.01' begins with a lowercase letter and is a fragment; it should be integrated into the preceding sentence.","section":"Section 4.1"},{"comment":"The post-processing section states that 'we select the smaller mode to favour parsimony' when the posterior is multimodal, but it does not describe how the two methods' estimates are reconciled when they disagree; a sentence explaining the practical consequence for the reported G+ would be helpful.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of stat.ME and the model is a reasonable contribution. The main concern is not correctness of the derivations but the strength of the empirical evidence for the central co-clustering claim; the shared-latent-space assumption is never varied in simulation, and the Aarhus application relies on post-hoc exclusions and an interpretability override. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also encourage the editor to ask for the R code verification of the Aarhus and primary-school analyses, since reproducibility is a stated strength of the submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"LaPCoM is worth a look. It is a genuine new combination, a two-level mixture-of-mixtures that clusters networks and nodes simultaneously under a latent position model, with MFM priors and automatic selection of the number of clusters. The paper is well specified, the supplement includes full conditionals and MCMC pseudocode, the PPCs are extensive, and the authors are admirably honest about limitations (fixed dimension, no temporal dependence, Poisson underfitting counts, no directionality). That honesty is earned: the simulations and applications mostly support the model.\n\nThe soft spots are real but not disqualifying. The most important is the shared-latent-space assumption: every network in a network-level cluster contributes the same Z_g. If the networks in a true cluster have heterogeneous latent structure, the model will either split the cluster or estimate a compromise latent space that blurs node-level clusters. The simulations never stress this, since they generate data from the same model, so the strong recovery results are self-consistency checks. The Aarhus application gives a hint: within the four-network cluster, per-network LPCM fits show 1, 3, 7, and 1 node clusters, which suggests heterogeneity that LaPCoM's shared Z_1 cannot represent. This deserves testing in revision, for example with network-specific latent perturbations or sender-receiver effects.\n\nSecond, the Aarhus analysis contains two post-hoc decisions: two chains are dropped because a K=5 node-cluster solution was called 'spurious', and the post-processing result (which merged the Facebook cluster into one) is overridden because K=2 is considered 'more interpretable'. These are disclosed, but they weaken the real-data demonstration and should be better justified.\n\nThird, PopNet beats LaPCoM on network-level ARI in every simulated comparison. LaPCoM's counterarguments, such as node-level clustering, count-valued support, and faster runtime, are valid, but the reader should see this in context: network-level clustering is not uniformly superior.\n\nThe count-valued primary-school application underfits large edge counts, and the authors admit the Poisson is overly restrictive. Minor, given the clustering focus.\n\nOverall, this is a serious methods paper for statisticians working on multiplex networks. I would send it to peer review, with a request that the authors address the shared-latent-space robustness and clean up the post-hoc selection in Aarhus. The code is public, the derivations are complete, and the model is genuinely useful. I would cite it if I were working on network clustering.","headline":"A useful new co-clustering model for multiplex networks, well-built but with an untested shared-latent-space assumption and a couple of post-hoc real-data decisions that need fixing.","tokens_in":40557,"tokens_out":2828,"would_cite":true,"duration_ms":32168,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62F15","62P25"],"pacs":[],"model":"deepseek-v4-flash","headline":"LaPCoM shows that a hierarchical latent position model can co-cluster the networks of a multiplex and the nodes within each network cluster, with cluster counts inferred from the data.","keywords":["multiplex networks","latent position model","co-clustering","mixture of finite mixtures","Bayesian nonparametrics","model-based clustering","community detection","Markov chain Monte Carlo"],"falsifier":"Generate a two-cluster multiplex in which the two clusters have identical latent positions and node communities but different per-network intercepts, or a within-cluster node-specific position shift, then fit LaPCoM; consistently recovering more than two network clusters, or node partitions that differ across networks in the same cluster, would falsify the shared-$Z_g$ exchangeability assumption.","tokens_in":39522,"feed_emoji":"🕸️","tokens_out":10875,"duration_ms":104427,"temperature":0.7,"pith_summary":"LaPCoM is a Bayesian model that treats a multiplex as a collection of networks generated from a two-level mixture: first networks are grouped into clusters that share a latent position space, and then, inside each such space, nodes are grouped by a Gaussian mixture on their latent positions. The paper's central claim is that this hierarchical latent-space formulation lets one recover both levels of structure at once, with the number of clusters at each level learned from the data through a mixture-of-finite-mixtures prior rather than fixed in advance. If the claim holds, practitioners gain a single interpretable tool for multiplex data that currently requires separate network-clustering and node-clustering analyses, and it works for both binary and count-valued edges. Simulations show accurate recovery of latent spaces and clusterings, and the real-data applications yield groupings with substantive meaning, such as lesson versus lunchtime network regimes and class-based node communities.","feed_headline":"One model clusters multiplex networks and their nodes together","feed_subtitle":"LaPCoM learns shared latent maps for network groups and node communities, choosing cluster counts from the data.","key_machinery":"The load-bearing object is the hierarchical mixture-of-mixtures latent position model: a top-level mixture over networks, each component carrying a shared latent position matrix $Z_g$, and a bottom-level Gaussian mixture on the rows of $Z_g$. A dynamic mixture of finite mixtures prior, with translated beta-negative-binomial priors on $G$ and each $K_g$ and Dirichlet concentration parameters scaled by the component count, is what lets the model empty superfluous components and choose cluster numbers automatically; telescoping sampling updates the component counts, and Procrustes alignment plus label-permutation post-processing resolves rotational and label-switching identifiability.","core_discovery":"On its own terms, the paper establishes that a multiplex can be modelled as $Y^{(m)} \\sim \\sum_{g=1}^{G} \\tau_g \\prod_{i\\ne j} P(\\lambda_{g,ij})$, with $f(\\lambda_{g,ij}) = \\alpha - \\|z_{g,i} - z_{g,j}\\|_2^2$, where each network-level component $g$ has its own latent position matrix $Z_g$; the rows of $Z_g$ are then drawn from a Gaussian mixture $\\sum_{k=1}^{K_g} \\pi_{gk} \\mathrm{MVN}_2(\\mu_{gk}, \\Sigma_{gk})$, which induces node clusters within that network cluster. The number of network components $G$ and node components $K_g$ are treated as random under translated $\\beta$-negative-binomial priors, with Dirichlet shrinkage on the mixing weights, so the fitted model selects active clusters automatically; inference uses a Metropolis-within-Gibbs sampler with telescoping updates. Simulation studies report near-perfect or high adjusted Rand indices at both levels, and the model matches or beats comparison methods on network clustering while providing node-level clustering that they lack. Applications to three multiplexes yield network-level clusters aligned with context, such as perceived advice density, Facebook versus other ties, and lunchtime versus class hours, and node-level clusters aligned with academic roles or school classes.","pith_inferences":["A natural boundary of the shared-$Z_g$ assumption is that network-specific sender, receiver, or density effects would require either more network clusters or an extended model with per-network intercepts or random latent shifts; this is testable by simulation and is a likely next step.","The primary school application suggests that replacing the Poisson edge distribution with a zero-inflated or overdispersed count model would improve fit on weighted multiplexes without changing the co-clustering structure; the paper itself flags the Poisson restriction.","The fixed two-dimensional latent space could be relaxed by shrinkage priors on the latent dimension, which would let the model decide whether a network cluster needs more than two dimensions for its map.","The co-clustering output doubles as an exploratory visualization tool: each network-level cluster gets its own two-dimensional plot, and the node-level clusters inside it read directly as roles or communities, which could guide hypothesis generation on real multiplex data."],"forward_implications":["A single fitted model yields both a low-dimensional map of each network cluster and community labels for nodes inside it, so practitioners do not need separate network-level and node-level clustering pipelines.","The number of network clusters and of node clusters per network cluster is estimated from the data, so users are not forced to pre-specify $G$ and $K_g$.","Because the edge distribution is plug-in (Bernoulli for binary, Poisson for counts), the same co-clustering machinery applies to presence-absence and interaction-count multiplexes.","The shared latent space per network cluster offers a parsimonious alternative to fitting one latent space per network, with posterior predictive checks indicating comparable or better fit in the applications shown.","Network-level clusters can capture meaningful context regimes in time-stamped multiplexes, such as lesson versus lunchtime interaction periods, even without an explicit temporal model."],"supporting_citations":[{"why":"Defines the latent position model that LaPCoM generalises; the edge probability depends on squared Euclidean distance between latent positions.","marker":"Hoff et al. (2002)"},{"why":"Adds the Gaussian mixture on latent positions (LPCM) that LaPCoM uses for node-level clustering inside each network cluster.","marker":"Handcock et al. (2007)"},{"why":"Supplies the dynamic mixture of finite mixtures, beta-negative-binomial priors on component counts, and telescoping sampler that give automatic cluster-number selection.","marker":"Fr¨ uhwirth-Schnatter et al. (2021)"},{"why":"Provides the label-switching post-processing procedure adopted at both levels and a multiplex node-clustering benchmark.","marker":"D'Angelo et al. (2023)"},{"why":"PopNet is the main network-clustering comparison method in the simulation study.","marker":"Durante et al. (2017)"},{"why":"Mixture of measurement-error network models used as a competing method for network and node clustering.","marker":"Mantziou et al. (2024)"},{"why":"latentnet per-network LPCM fits serve as the baseline for posterior predictive fit comparisons in the applications.","marker":"Krivitsky and Handcock (2008)"}],"fun_headline_variants":["Unified model co-clusters multiplex networks and their nodes","Latent positions give joint network and node clusters in multiplexes","Bayesian model co-clusters networks and nodes in multiplex data","One model finds network groups and node communities together"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Networks assigned to the same network-level cluster are assumed to share one latent position matrix, so all differences between those networks, such as density, sender or receiver tendencies, or temporal drift, must be absorbed by the common intercept and the shared positions; if real same-cluster networks differ in latent positions, the model will either split the cluster or distort the shared space and the node clusters with it.","fun_headline_variants_meta":{"raw":{"variants":["Unified model co-clusters multiplex networks and their nodes","Latent positions give joint network and node clusters in multiplexes","Bayesian model co-clusters networks and nodes in multiplex data","One model finds network groups and node communities together"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1897,"prompt_tokens":1063,"completion_tokens":834,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":767}},"tokens_in":679,"tokens_out":834,"duration_ms":7235,"temperature":1.0,"reasoning_tokens":767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:58:05.161724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a two-cluster multiplex in which the two clusters have identical latent positions and node communities but different per-network intercepts, or a within-cluster node-specific position shift, then fit LaPCoM; consistently recovering more than two network clusters, or node partitions that differ across networks in the same cluster, would falsify the shared-$Z_g$ exchangeability assumption.","supporting_citations":[{"cited_title":"S., Raftery, A","cited_arxiv_id":null,"evidence_quote":"Adds the Gaussian mixture on latent positions (LPCM) that LaPCoM uses for node-level clustering inside each network cluster."},{"cited_title":"u hwirth-Schnatter, S., Malsiner-Walli, G., and Gr \\","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic mixture of finite mixtures, beta-negative-binomial priors on component counts, and telescoping sampler that give automatic cluster-number selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the label-switching post-processing procedure adopted at both levels and a multiplex node-clustering benchmark."},{"cited_title":"B., and Vogelstein, J","cited_arxiv_id":null,"evidence_quote":"PopNet is the main network-clustering comparison method in the simulation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mixture of measurement-error network models used as a competing method for network and node clustering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"latentnet per-network LPCM fits serve as the baseline for posterior predictive fit comparisons in the applications."}],"review_version":1}