{"id":"45221676-ffcf-46f0-bcd8-572deb7002b9","arxiv_id":"2412.16298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An iterative profile least squares algorithm jointly estimates covariate coefficients and latent positions in networks, with bootstrap-based inference.","lead":"This paper develops an iterative profile least squares algorithm for network models where edge probabilities depend on observed covariates plus an unobserved low-rank latent structure. The method separates covariate-driven from latent-driven network structure and provides bootstrap confidence intervals for covariate effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The alternating estimator has no identifiability or consistency guarantee when covariates correlate with latent structure; the simulations always generate covariates independently, so the central empirical claim is untested in exactly the confounding regime.","rationale":"Read in good faith, this is a method paper: it proposes a practical iterative profile least squares algorithm and supports it with extensive simulations, a comparison to Mu et al. (2022) in Section 5.1, and four real-data illustrations. The simulation protocol, bootstrap coverage checks, and reproducible comparisons are real evidence and should be credited. My concern targets the boundary of the central claim, not the implementation. Section 3.1 is explicit that the spectral filter F is nonlinear, so the classical profile least squares closed form does not apply; the iterative replacement is heuristic unless a fixed-point or consistency argument is supplied. The paper supplies none: no proposition states consistency of gamma_hat, no identifiability condition rules out overlap between X gamma and Lambda I_{qs} Lambda^T, and no regularity condition addresses covariate-latent dependence. This matters because Section 6 interprets the estimated decomposition substantively, for example by attributing alliance structure to trade and democracy. Those interpretations presuppose that the decomposition is identified and that the alternating algorithm finds the identified pair. The simulations in Section 5 always generate covariates independently of the latent block labels, so they cannot detect confounding or non-identifiability. The reader's weakest assumption identifies essentially this same gap, and my proposed simulation makes it operational. If the algorithm degrades under correlated designs, the conditional verdict is warranted and the paper should state the independence condition explicitly; if it survives, the concern is minor and the evidence for the central claim is stronger. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":22580,"tokens_out":4282,"duration_ms":43109,"concrete_test":"Simulate the Type I model from Section 5.1 with an endogenous covariate: draw block labels z_i and latent positions alpha_i = mu_{z_i}, then generate node covariate xi_i = beta * alpha_i + epsilon_i with epsilon_i ~ N(0,1), and set the edge covariate x_ij = |xi_i - xi_j|. Sweep beta over 0, 0.5, 1, and 2 with the true gamma fixed. Run Algorithm 1 with the paper's settings (T = 20 initializations, M = 500 iterations) over 250 replications, and compute MSE(gamma_hat), mean ARI, and 95% bootstrap coverage of gamma. Also initialize once from the true gamma to separate local-optimum failure from identifiability failure. If MSE or coverage degrades as beta increases, the central claim is conditional on covariate-latent independence; if performance stays flat, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1's Algorithm 1 alternates Eq. (3.6), the OLS update of gamma after subtracting the current low-rank estimate, with Eq. (3.3), the spectral embedding of Y(gamma) = A - X gamma. The paper explicitly notes that the spectral filter F is nonlinear (Section 3) and provides no theorem that this fixed-point iteration recovers (gamma, Lambda), no identifiability condition separating X gamma from Lambda I_{qs} Lambda^T, and no consistency result for the estimated gamma used inside the spectral step. This is not merely a missing proof: the objective is nonconvex, and the spectral step is only justified when the matrix being embedded is true low-rank plus noise. If covariates are correlated with latent positions, as is plausible for trade and democracy with alliance structure in Section 6, then A - X gamma_hat can have a distorted low-rank signal, and the OLS update may absorb part of the latent structure into gamma_hat. The simulation designs in Section 5 draw covariates independently of z_i and alpha_i, so they never exercise this confounding regime. Consequently, the paper's central claims of accurate gamma estimation and cluster recovery are conditional on an unstated covariate-latent independence assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a single undirected binary network model in which the edge probability matrix is P_ij = x_ij^T gamma + alpha_i^T I_qs alpha_j, a linear edge-covariate term plus an indefinite-inner-product latent-position term. It proposes an iterative profile least squares estimator in which the residual matrix A - X gamma is spectrally embedded and gamma is updated by OLS on the resulting residuals (Algorithm 1), and it develops bootstrap inference based on the generalized bootstrap of Chatterjee and Bose (2005) with Bayesian bootstrap weights and GMM clustering of the estimated latent positions. The method is evaluated on simulations with SBM-type and general low-rank residual structures and applied to tree, physician-friendship, and military-alliance networks.","tokens_in":22995,"tokens_out":9103,"duration_ms":85313,"significance":"If the algorithm and bootstrap procedure were theoretically supported, the paper would offer a useful practical decomposition of network structure into observed covariate effects and latent heterogeneity, with a computationally simple estimation recipe and bootstrap-based uncertainty quantification. The paper's genuine strengths are its explicit low-rank indefinite-kernel formulation, the breadth of simulation settings (categorical and continuous covariates; stochastic-blockmodel and indefinite low-rank residuals), the comparison with Mu et al. (2022) in the categorical case, and the four real-data illustrations. However, the central methodological claims are currently supported only by simulation and application: no identifiability, convergence, or consistency results are proved for the alternating procedure, and the simulation design does not cover covariate-latent dependence, which is precisely the regime in which the separation between X gamma and Lambda I_qs Lambda^T is most questionable.","major_comments":[{"comment":"The paper provides no identifiability condition separating the linear term X gamma from the latent term Lambda I_qs Lambda^T, and no convergence or consistency result for the alternating profile least squares procedure. This is a load-bearing gap because the spectral embedding step in Step 3 of Algorithm 1 is only justified when the matrix being embedded is the true low-rank residual plus noise; if covariates are correlated with latent positions, A - X hat_gamma can carry a distorted low-rank signal and the OLS update in Eq. (3.6) can absorb part of Lambda into hat_gamma. The simulation study in Section 5 generates covariates independently of the latent positions z_i and alpha_i, so it never exercises this confounding regime. I ask the authors either to state explicit identifiability and compatibility conditions and provide (or cite) theory for the alternating estimator under those conditions, or to weaken the paper's claims to a heuristic method and add a simulation setting in which covariates depend on the latent cluster labels (e.g., x_i drawn from a cluster-dependent distribution) to assess robustness.","section":"Section 3.1, Eqs. (3.3)-(3.6) and Algorithm 1"},{"comment":"The bootstrap inference is built on Chatterjee and Bose (2005), but no condition of that theory is verified for this nonconvex spectral-profile estimator, and the resampling step for Lambda is not shown to solve the corresponding resampled estimating equations. The de-weighting step alpha*_ib = alpha^w_ib / sqrt(W_bi) is motivated by a heuristic reading of Eq. (4.2) rather than derived from the weighted estimating equations. Moreover, 'correct coverage' is claimed on the basis of Figs. 4 and 8, which show confidence intervals from a single randomly generated network per setting; no coverage proportions over the 250 replications are reported. Please provide either the distributional conditions and coverage theory, or clearly label the bootstrap as heuristic and report empirical coverage rates over replications.","section":"Section 4, Eqs. (4.1)-(4.2), and Figs. 4/8"},{"comment":"The reported ARI improvement is obtained by manually re-running the MClust step with K=2 for 11 of 250 replications after the default pipeline selected K=3. This is a post-hoc modification of the clustering procedure after seeing the results, and it inflates the apparent cluster recovery of the default method. Please report both the default and the overridden ARI values, and specify in advance a reproducible rule for overriding the model-selection step (for example, a threshold on the ratio of MClust uncertainties, or a clearly stated visual-check criterion). Without such a rule the cluster-recovery results are not reproducible.","section":"Section 5.1, Fig. 3 and following text"}],"minor_comments":[{"comment":"The notation |Y_m| = (Y_m^T Y_m)^{1/2} combined with 'arranged in decreasing order (based on their actual, signed, value)' is ambiguous; clarify that the eigendecomposition of Y_m is used and that the d eigenvalues largest in absolute value are retained, with q_m and s_m equal to the numbers of positive and negative retained eigenvalues.","section":"Algorithm 1, Step 3"},{"comment":"The text refers to a '3 x 3 block-form' and to blocks (1,3), (2,3), and (3,3) even though only K=2 residual blocks were defined; please clarify whether three blocks are intended or correct the block indexing, as this affects the reproducibility of the simulation design.","section":"Section 5.2, settings (b) and (c)"},{"comment":"The statement that Exponential(alpha) weights with alpha=1 and alpha=O(n^{-1/2}) have performances comparable to the naive and m-out-of-n bootstraps is not accompanied by a reference or simulation evidence; please cite a source or qualify the statement.","section":"Section 4, bootstrap weights"},{"comment":"The captions state l=1,...,13 for the alliance data, but this dataset has p=6 covariates; update the captions to the correct index range.","section":"Figs. 16 and 18 captions"},{"comment":"The choice W_bi ~ Exponential(alpha) is ambiguous about the rate/scale parameterization; state the density or mean explicitly so that the reported alpha = n^{-1/2} is reproducible.","section":"Section 4, parameterization of Exponential weights"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned as a statistical methodology contribution but currently lacks identifiability and consistency theory for its headline estimator; if the journal's bar for stat.ME requires such theory, the issues in the major comments are blocking. If the journal welcomes applied methodology supported by extensive simulation, the reproducibility issues in Section 5.1 and the absence of a confounding simulation still need to be addressed. The reliance on Chatterjee and Bose (2005), which is co-authored by one of the present authors, is appropriate, but its conditions should not be assumed to transfer automatically to this nonconvex spectral-profile estimator."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nQuick take: this paper gives applied network researchers a simple iterative profile least squares routine for a GRDPG-with-covariates model, allowing continuous and mixed edge covariates and general low-rank residual structure. It is a natural extension of Mu et al. (2022) and Rubin-Delanchy et al. (2022), and the bootstrap inference (based on Chatterjee and Bose 2005, where one author is on this paper) is a genuinely useful addition. The writing is clear and honest—they explicitly note that the spectral filter is nonlinear, so the classical PLS closed-form does not carry over, and they present the algorithm as a heuristic fixed-point iteration rather than an estimator with a proof.\n\nWhat is good: the model is clearly defined, the Type I (SBM) vs Type II (general low-rank) distinction is helpful, and the simulations cover several covariate types and residual structures. The four real-data examples are analyzed with care—the decomposition into covariate effect and residual matrix is interpretable, and the bootstrap confidence intervals on the alliance networks agree with substantive findings in the literature. I also appreciate that they discuss the post-hoc override of K in 11/250 replications in Section 5.1 rather than burying it.\n\nThe soft spots are real, though. The central Algorithm 1 has no identifiability condition separating Xγ from Λ I_{qs} Λ^T and no consistency or convergence result; the theory ends at \"this seems to work.\" That alone is not disqualifying for an applied paper, but the bigger issue is that every simulation generates covariates independently of the latent positions. That never tests the confounding regime. The moment the covariates are correlated with the latent structure—which is precisely what you expect in the alliance and CKM data—the OLS update can absorb part of the latent structure into γ̂, and the spectral embedding step is then operating on a distorted signal. So the paper's central claim of accurate γ estimation and cluster recovery is conditional on an unstated independence assumption. This should have been acknowledged and ideally investigated in the simulations.\n\nThe Section 5.1 K override is a minor concern, not a fatal one: they disclose it, give a plausible reason (MClust overestimates K on a noisy residual), and show the uncertainty drops when K=2. But the ARI numbers in Fig 3(b) are therefore less clean than they appear.\n\nBottom line: this is a useful paper for applied network researchers, and it deserves a serious referee. But the editor should send it back for major revision. The authors need to either add some theoretical support (even for a simplified version), or substantially tone down the claims and add a simulation study where covariates are correlated with the latent positions—maybe just a few lines showing that the method breaks or degrades, so users know the operating range. I'd engage with it, but not cite it in its current form for anything beyond a heuristic comparison.\n\nRecommendation: send to peer review, with a clear expectation of heavy revision.\n\nCheers,\n[Your name]","headline":"Useful, clearly written extension of GRDPG-with-covariates to continuous and mixed covariates, but the alternating estimator has no convergence or consistency theory and the simulations never test the covariate-latent confounding regime that the real applications likely live in.","tokens_in":23349,"tokens_out":3715,"would_cite":false,"duration_ms":34063,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62F40","62H30","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes an iterative profile least squares algorithm that estimates covariate effects and latent network structure simultaneously by alternating spectral embedding of the covariate-adjusted adjacency matrix with linear…","keywords":["spectral estimation","homophily","generalized random dot product graphs","network visualization","profile least squares","bootstrap inference","stochastic blockmodels","edge covariates"],"falsifier":"Simulate networks under the paper's model with $d=2$, $q=s=1$, and with edge covariates generated as functions of the latent positions so that covariate and latent structure are correlated; if Algorithm 1's estimates of $\\gamma$ show bias that persists as $n$ grows, or if nominal 95% bootstrap intervals undercover substantially, the assumed separation of covariate and latent effects is not achieved.","tokens_in":22416,"feed_emoji":"🔗","tokens_out":9671,"duration_ms":81390,"temperature":0.7,"pith_summary":"The paper tries to establish that, in a single observed network, the part of link formation explained by observed node or edge attributes can be cleanly separated from the part driven by unobserved latent structure. Its model sets the probability of an edge to a linear function of edge covariates plus an indefinite inner product of latent node positions, a kernel flexible enough to represent both assortative and disassortative residual structure. Because the latent positions are unknown, the classical closed-form profile least squares estimator does not transfer; the paper replaces it with an iteration that spectrally embeds the covariate-adjusted adjacency matrix and then updates the covariate coefficients by least squares. Simulations with stochastic blockmodel residuals and with general low-rank residuals show mean squared error of the coefficients decreasing with network size, and Bayesian bootstrap confidence intervals are shown to contain the true parameter values. Applications to a tree–fungus network, a physician friendship network, and two military-alliance networks illustrate how the covariate and residual contributions can be visualized and tested.","feed_headline":"One algorithm splits network links into observed and hidden causes","feed_subtitle":"Finds which covariates drive network links, then isolates the residual structure left for latent factors.","key_machinery":"The central object is the indefinite inner product kernel $f(x,y;q,s)=\\sum_{l=1}^{q} x_l y_l - \\sum_{l=q+1}^{q+s} x_l y_l$, which lets the residual part of the model capture both homophily (first $q$ dimensions) and heterophily (last $s$ dimensions). The argument is carried by an alternating scheme: for fixed $\\gamma$, spectral embedding of $Y(\\gamma)=A-X\\gamma$ through the eigendecomposition retaining the $d$ largest eigenvalues in signed magnitude estimates the residual kernel, with the latent dimension chosen by a profile-likelihood scree rule; for that kernel estimate, the update $\\hat\\gamma = (\\tilde X^T\\tilde X)^{-1}\\tilde X^T \\mathrm{vec}(A - \\Lambda I_{qs}\\Lambda^T)$ is the usual least squares solution. A Bayesian bootstrap that assigns random weights to the estimating equations supplies confidence intervals for $\\gamma$ and for the residual intensities.","core_discovery":"The central claim is that the covariate coefficient vector and the latent node positions can be estimated jointly by profile least squares even though the projection filter used to embed the residual matrix is nonlinear, so no closed-form estimator exists. For a fixed coefficient vector $\\gamma$, the matrix $Y(\\gamma)=A-X\\gamma$ is treated as a generalized random dot product graph adjacency matrix: its top signed eigenvalues and eigenvectors give an estimate of the residual kernel $\\Lambda I_{qs}\\Lambda^T$. For this estimated kernel, the coefficient vector updates by ordinary least squares regression of $A$ minus the estimated kernel on the edge covariates. The paper argues that alternating these two steps to convergence yields estimates of $\\gamma$ and of the residual structure, with simulations demonstrating the behavior of the estimates as $n$ grows and with the bootstrap providing uncertainty statements.","pith_inferences":["Editorial extension: the same profile least squares scheme could be run on directed or weighted networks by replacing the symmetric GRDPG embedding with a directed or weighted spectral embedding; the estimating equations for $\\gamma$ would remain least squares.","Editorial extension: a stress test the paper does not run is to generate covariates as functions of the latent positions, so that covariate and latent structure are correlated, and to track bias in $\\hat\\gamma$; the alternating procedure's identifiability in that regime is the main open question.","Editorial extension: because the type II residual matrix is not a valid probability matrix, the method effectively functions as a matrix-decomposition diagnostic; applying it to networks with known ground-truth covariate effects could benchmark how much of the latent structure the covariate term absorbs.","Editorial extension: a formal consistency proof for the alternating estimator would be the natural next step; the simulation evidence suggests the separation works in the regimes tested, but no theorem yet guarantees it."],"forward_implications":["Users of the method can decompose the estimated edge probability matrix into a covariate effect matrix $x_{ij}^T\\hat\\gamma$ and a residual kernel estimate, so the visual contribution of observed attributes versus unobserved factors is directly inspectable.","The method applies to continuous, categorical, and mixed covariates, covering cases where earlier spectral community-detection approaches with vertex covariates only handle discrete attributes.","Bayesian bootstrap weights on the estimating equations yield confidence intervals for $\\gamma$ and for residual intensities, enabling significance statements such as geographic distance in the tree network, city of practice in the physician network, and trade and joint democracy in the alliance networks.","Simulations under both block-model residuals and general low-rank residuals show that mean squared error of $\\hat\\gamma$ decreases and adjusted Rand index increases with network size, and the displayed bootstrap intervals contain the true parameters."],"supporting_citations":[{"why":"Provides the indefinite inner product kernel definition and the spectral embedding result that justifies estimating latent positions from the covariate-adjusted adjacency matrix.","marker":"Rubin-Delanchy et al. (2022)"},{"why":"Supplies the discrete-covariate stochastic blockmodel that the paper compares against; with a single binary covariate, the paper's model reduces to theirs.","marker":"Mu et al. (2022)"},{"why":"Establishes the generalized bootstrap for estimating equations, which the paper adapts with Bayesian bootstrap weights for inference.","marker":"Chatterjee and Bose (2005)"},{"why":"Defines classical profile least squares in partial linear models, the template whose closed-form estimator fails here because the spectral filter is nonlinear.","marker":"Speckman (1988)"},{"why":"Supplies the random dot product graph spectral embedding background that motivates using eigenvectors of the residual matrix as latent position estimates.","marker":"Athreya et al. (2017)"},{"why":"Provides the profile-likelihood scree rule used in Algorithm 1 to choose the latent dimension.","marker":"Zhu and Ghodsi (2006)"},{"why":"Supplies the tree network data and the logistic regression goodness-of-fit framework whose findings on covariate significance the tree example is checked against.","marker":"Latouche et al. (2018)"},{"why":"Documents the percentile and basic bootstrap confidence interval constructions used for the simulation and data example intervals.","marker":"Davison and Hinkley (1997)"}],"fun_headline_variants":["Profile LS splits network ties into explained and hidden","New method separates covariate effects from latent structure","Algorithm isolates observed and unobserved network drivers","Jointly estimate covariates and residual network kernel","Profile least squares uncovers latent vs observed link drivers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that at each iteration the covariate-adjusted residual matrix $A-X\\gamma$ is close to a low-rank symmetric matrix whose top eigenvectors consistently estimate the true latent positions, even while $\\gamma$ is being estimated and even if the covariates are correlated with the latent structure.","fun_headline_variants_meta":{"raw":{"variants":["Profile LS splits network ties into explained and hidden","New method separates covariate effects from latent structure","Algorithm isolates observed and unobserved network drivers","Jointly estimate covariates and residual network kernel","Profile least squares uncovers latent vs observed link drivers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1182,"prompt_tokens":920,"completion_tokens":262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":192}},"tokens_in":536,"tokens_out":262,"duration_ms":3172,"temperature":1.0,"reasoning_tokens":192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:42:59.436195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate networks under the paper's model with $d=2$, $q=s=1$, and with edge covariates generated as functions of the latent positions so that covariate and latent structure are correlated; if Algorithm 1's estimates of $\\gamma$ show bias that persists as $n$ grows, or if nominal 95% bootstrap intervals undercover substantially, the assumed separation of covariate and latent effects is not achieved.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the indefinite inner product kernel definition and the spectral embedding result that justifies estimating latent positions from the covariate-adjusted adjacency matrix."},{"cited_title":"and Bose, A","cited_arxiv_id":null,"evidence_quote":"Establishes the generalized bootstrap for estimating equations, which the paper adapts with Bayesian bootstrap weights for inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines classical profile least squares in partial linear models, the template whose closed-form estimator fails here because the spectral filter is nonlinear."},{"cited_title":"E., Tang, M., Priebe, C","cited_arxiv_id":null,"evidence_quote":"Supplies the random dot product graph spectral embedding background that motivates using eigenvectors of the residual matrix as latent position estimates."},{"cited_title":"and Ghodsi, A","cited_arxiv_id":null,"evidence_quote":"Provides the profile-likelihood scree rule used in Algorithm 1 to choose the latent dimension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the tree network data and the logistic regression goodness-of-fit framework whose findings on covariate significance the tree example is checked against."}],"review_version":1}