{"id":"c4cbf4e1-6a5b-41a7-a3e7-b53f797ed03d","arxiv_id":"2505.02986","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A covariate-assisted latent space model with horseshoe shrinkage on both covariate loadings and network-covariate mismatches achieves adaptive posterior contraction and improved community detection rates.","lead":"This paper proposes a Bayesian model that learns network communities while automatically ignoring useless or misleading node covariates, by letting each node decide how much covariate information to trust. The method comes with fast variational algorithms and a theoretical guarantee that covariate information improves community detection when it is relevant and does no harm when it is not.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3's proof drops a √n factor: the derived misclustering rate should be n e^{−β*} ε_n^2 / δ^2, not e^{−β*} ε_n^2 / δ^2, so the claimed improvement over network-only clustering is not established.","rationale":"The reader's projection concern about Assumption 6 is legitimate and self-admitted in Section 3: the community-detection theorem is proved for a projected prior that the algorithm does not implement, so the method as run is not formally covered. However, the more acute problem is internal to the proof of Theorem 3.3. The final step of Appendix A.4 drops a factor of √n: Corollary 3.2 bounds the Gram-matrix Frobenius error by n e^{−β*/2} ε_n, and Lemma A.6 divides by √n, yielding √n e^{−β*/2} ε_n for the Procrustes error, not e^{−β*/2} ε_n. Consequently Lemma A.7 gives a misclustering rate that is n times larger than stated. The theorem's own assumption (4) contains the same extra n, confirming the inconsistency. The corrected rate matches, rather than improves, the network-only rate under the usual scaling log p = O(log n), so the paper's most distinctive theoretical promise is unsupported. This is not a matter of disagreeing with a consensus; it is a checkable algebraic inconsistency within the paper's own equations. The projection gap is real but would be secondary even if the scaling error were repaired. A conditional accept might be appropriate if the authors could supply a missing argument that recovers the extra √n, but the proof as written does not do so, and the central claim is the improved misclustering rate. I therefore recommend REJECT until the theorem and proof are reconciled.","tokens_in":27109,"tokens_out":13288,"duration_ms":134494,"concrete_test":"Re-derive the norm chain in Appendix A.4 without dropping factors: take the high-probability event from Corollary 3.2, compute ||X̂ X̂' − X* X*'||_F, apply Lemma A.6 with σ_d(X*) ≍ √n, then Lemma A.7, and compare the final misclustering rate with the stated one. If the extra n persists, either rerun the proof with a (1/n) Σ normalization in Corollary 3.2 or check whether Theorem 3.3's minimal block-size condition should be ω(n^2 e^{−β*} ε_n^2 / δ^2); either way the claimed comparison with the network-only rate e^{−β*} d log n/(n δ^2) must be recomputed.","verdict_should_be":"REJECT","load_bearing_attack":"Corollary 3.2 controls (1/n^2) Σ_{i,j}(x_i'x_j − x*_i'x*_j)^2 ≤ M e^{−β*} ε_n^2 with high fractional-posterior probability, so ||X̂ X̂' − X* X*'||_F ≤ M^{1/2} n e^{−β*/2} ε_n. Appendix A.4 then invokes Lemma A.6 with σ_d(X*) ≍ √n to conclude ||X̂ − X* O||_F ≲ ||X̂ X̂' − X* X*'||_F / √n ≲ e^{−β*/2} ε_n. That last step drops the factor n from the Corollary bound; the correct conclusion is ||X̂ − X* O||_F ≲ √n e^{−β*/2} ε_n. Feeding this into Lemma A.7 gives L(Ξ̂, Ξ*) ≲ ||X̂ − X* O||_F^2 / δ^2 ≲ n e^{−β*} ε_n^2 / δ^2, not e^{−β*} ε_n^2 / δ^2. This is internally visible in the theorem's own block-size condition (4), which uses ω(n ε_n^2 e^{−β*} / δ^2), matching the corrected rate. Since ε_n^2 = (1+s_b+s_x)d log(np)/n^2, the corrected misclustering rate is e^{−β*}(1+s_b+s_x)d log(np)/(n δ^2), which is the same order as the network-only rate e^{−β*} d log n/(n δ^2) when p is polynomial and s_b+s_x are bounded, and worse when those constants grow. The advertised adaptive improvement in community detection therefore rests on a scaling error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Bayesian latent space model (CALSM) for network data with high-dimensional node covariates. The model posits X approximately equal to ZB with B row-sparse, and uses horseshoe-type shrinkage on both the covariate coefficients and the node-level deviations from ZB. The authors develop two variational inference algorithms (CAVI with an R/Rcpp implementation, and SVI with edge subsampling for large sparse networks) and analyze the fractional posterior in the homogeneous sparse network setting. The main theoretical results are Theorem 3.1, giving contraction of the fractional posterior for the link-probability matrix in average alpha-Renyi divergence at rate epsilon_n^2 = M_0^2(1+s_b+s_x)d log(np)/n^2; Corollary 3.2, controlling the average squared error of inner products with an e^{-beta*} factor; and Theorem 3.3, which states a misclustering rate after K-means of L(Xihat, X*) less than or similar to e^{-beta*} epsilon_n^2/delta^2 and claims an improvement over the network-only rate e^{-beta*} d log n/(n delta^2). Simulations and three real-world datasets, plus a large PubMed example, are used to demonstrate estimation and clustering performance.","tokens_in":27543,"tokens_out":9224,"duration_ms":88703,"significance":"The paper addresses a relevant problem, and the general strategy of using shrinkage priors on both covariate coefficients and node-level discrepancies is natural. If the stated rates were correct, the model would deliver a principled way to adaptively combine network and covariate information, degrading gracefully when covariates are uninformative. The manuscript is transparent: it explicitly states that Assumption 6 (a projection to bounded and centered latent vectors) is used only for the proofs and is not used in the implemented algorithms, and it provides detailed appendices with prior-concentration bounds and algorithm derivations. The principal importance of the paper rests on the community-detection improvement claimed in Theorem 3.3, and that claim is not supported by the proof as written.","major_comments":[{"comment":"The proof of Theorem 3.3 drops a factor of n when converting the inner-product bound of Corollary 3.2 into a bound on the latent positions. Corollary 3.2 controls (1/n^2) sum_{i,j} (x_i'x_j - x_i*'x_j*)^2 <= M e^{-beta*} epsilon_n^2, hence ||X X' - X* X*'||_F less than or similar to n e^{-beta*/2} epsilon_n. Lemma A.6 then yields ||X - X*O||_F less than or similar to sqrt(n) e^{-beta*/2} epsilon_n, not e^{-beta*/2} epsilon_n as written in Appendix A.4. Plugging this into Lemma A.7 gives L(Xihat, X*) less than or similar to n e^{-beta*} epsilon_n^2/delta^2, matching the block-size condition (4) but contradicting the displayed rate e^{-beta*} epsilon_n^2/delta^2 in Theorem 3.3. Since epsilon_n^2 = (1+s_b+s_x)d log(np)/n^2, the corrected misclustering rate is e^{-beta*}(1+s_b+s_x)d log(np)/(n delta^2), which is the same order as the network-only rate e^{-beta*} d log n/(n delta^2) when p is polynomial and s_b+s_x are bounded, and worse when these constants grow. The advertised adaptive improvement in community detection is therefore not established.","section":"Theorem 3.3 / Appendix A.4"},{"comment":"Corollary 3.2 and Theorem 3.3 are proved under Assumption 6, which replaces the prior by its restriction to the bounded and centered event B_p(X). Section 3 explicitly states that this projection 'is not utilized in the algorithm,' and the claim that the projection has 'negligible difference' is asserted without proof. The contraction and misclustering guarantees therefore do not formally cover the CAVI or SVI posteriors actually computed. To close this gap, the paper needs a quantitative statement (e.g., a lower bound on Pi(B_p) or a uniform bound showing the variational ELBO is asymptotically unaffected by the projection) rather than an appeal to a remark in Ma et al. (2020).","section":"Section 3, Assumption 6"},{"comment":"Theorem 3.3 says 'Suppose the assumptions in Theorem 3.2 hold,' but there is no Theorem 3.2 in the manuscript. The evident intended statement is Corollary 3.2 (or a separate latent-vector convergence theorem). The theorem numbering should be fixed and all assumptions used in Theorem 3.3 should be stated explicitly.","section":"Section 3, Theorem 3.3"},{"comment":"The definitions of tau*_x in (A.7) and tau*_b in (A.13), and the inequalities following them, are difficult to verify: the quantities delta and tau* appear to have different implied scales (a factor of n appears in some denominators and not in others), and the constants K, K', K'' in (A.8)-(A.10) are not tracked through the final exponent. Because the claimed prior-mass lower bound e^{-K n^2 epsilon_n^2} is the basis of Theorem 3.1, these estimates need to be rechecked carefully.","section":"Appendix A.2, equations (A.7) and (A.13)"}],"minor_comments":[{"comment":"The inequality in Corollary 3.2 is written as '>= MD + 3eta/(1-alpha) e^{-beta*} epsilon_n^2', which should be '>= M(D+3eta)/(1-alpha) e^{-beta*} epsilon_n^2'; the missing parentheses make the statement ambiguous.","section":"Corollary 3.2 display"},{"comment":"The prior for x_i should specify the covariance as lambda_{xi}^2 tau_x^2 I_d rather than just lambda_{xi}^2 tau_x^2 for dimensional consistency.","section":"Section 2.1, equation (3)"},{"comment":"The caption of Figure 2 says 'SVD performs the second best' in weak-signal settings, while the text says SVDyzO is second best; this inconsistency should be resolved.","section":"Figure 2 caption and Section 5.1"},{"comment":"There are typesetting and terminology issues: 'Asumption 5' and 'indetifibility' should be corrected, and Lemma A.5 uses h^2 without defining it; the standard Hellinger affinity should be defined.","section":"Section 3 and Lemma A.5"}],"recommendation":"major_revision","confidential_remarks":"The sqrt(n) scaling error in Theorem 3.3 is the key obstacle. If the authors can prove a nontrivial improvement under a different regime (for example, p growing faster than polynomial in n, or s_b+s_x vanishing with n), the community-detection claim could be restored. The manuscript would also benefit from a careful re-verification of the prior-mass calculations in Appendix A.2. The reference list includes several self-citations (Zhao et al. 2022a,b; Niu et al. 2023) that are germane but should be checked for necessity. The missing Theorem 3.2 numbering should be fixed before further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does real new work: row-sparse covariate loadings plus node-specific horseshoe mismatch in a latent space model, with two variational algorithms and a fractional posterior contraction analysis. That combination is genuinely new relative to Zhang et al. (2022b) and spectral covariate-assisted methods. The simulations and real data are extensive, and the CAVI/SVI contributions are practical. Credit where due: the link-probability contraction theorem, if it checks out, is a solid contribution.\n\nBut the main advertised gain—faster community detection from covariates—has a scaling error in the proof. Corollary 3.2 controls (1/n^2) sum (x_i'x_j - x*_i'x*_j)^2 by M e^{-beta*} epsilon_n^2, so ||Xhat Xhat' - X* X*'||_F is at most n M^{1/2} e^{-beta*/2} epsilon_n. Appendix A.4 then divides by sigma_d(X*) ~ sqrt(n) and drops the n. The correct Procrustes bound is ||Xhat - X*O||_F ≲ sqrt(n) e^{-beta*/2} epsilon_n, which gives a misclustering rate n e^{-beta*} epsilon_n^2 / delta^2, not e^{-beta*} epsilon_n^2 / delta^2. Plugging in epsilon_n^2 = (1+s_b+s_x) d log(np)/n^2, that is the same order as the network-only rate when p is polynomial and s_b+s_x are bounded, so the claimed adaptive improvement in clustering is not established. The paper's own assumption (4) in Theorem 3.3 uses n epsilon_n^2 e^{-beta*}/delta^2, matching the corrected rate—so this is an internal inconsistency rather than a different convention. It needs to be fixed before the theorem is believable.\n\nSecond soft spot: Theorem 3.3 is proved for the projected prior in Assumption 6, which the algorithms do not implement. The authors flag this and point to a similar remark in Ma et al. (2020), but there is no argument that the projection is benign in this setting. If the rate claim is repaired, the theorem still does not formally cover the shipped CAVI/SVI posterior. The empirical results suggest the method works, but the theory is currently about a different estimator.\n\nMinor: the simulation figures lack error bars, and despite the supplement saying reproducible examples are attached, I do not see a code or data repository. That should be required for a methods paper.\n\nBottom line: the model and algorithms deserve referee time, and the contraction result is plausible, but the headline community-detection claim currently overstates. I would not cite the misclustering rate as is. Accept for peer review with major revision expected.","headline":"The method is genuinely new and worth a serious look, but the advertised community-detection improvement rests on a dropped sqrt(n) in the proof of Theorem 3.3.","tokens_in":28037,"tokens_out":2544,"would_cite":false,"duration_ms":25266,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A latent-space network model with horseshoe priors on covariate coefficients and node mismatches contracts at rate (1+s_b+s_x)d log(np)/n^2 and improves misclustering whenever a sparse covariate signal exists.","keywords":["network analysis","community detection","high-dimensional covariates","latent space model","horseshoe prior","fractional posterior","variational inference","sparse networks"],"falsifier":"Generate networks from the model with true latent vectors that violate the boundedness-and-centering event, for example by letting their norms grow with $n$ or giving them a nonzero row mean, then fit the published CAVI/SVI algorithms and measure misclustering as $n$ grows; if the error does not track $e^{-\\beta^*}(1+s_b+s_x)d\\log(np)/(n^2\\delta^2)$ but instead tracks the network-only $e^{-\\beta^*}d\\log n/(n\\delta^2)$, the projection assumption is load-bearing.","tokens_in":26896,"feed_emoji":"🕸️","tokens_out":9808,"duration_ms":100108,"temperature":0.7,"pith_summary":"This paper proposes a Bayesian latent-space model for networks with high-dimensional node covariates and claims that putting horseshoe shrinkage priors on two layers—the covariate-to-latent coefficient rows and the node-level discrepancies between network and covariate information—lets the posterior borrow covariate strength only where it is useful. The central theoretical claim is that the fractional posterior for link probabilities contracts at rate $\\epsilon_n^2 = M_0^2(1+s_b+s_x) d \\log(np)/n^2$ in average $\\alpha$-Rényi divergence, and that $K$-means on the estimated latent vectors misclusters at rate of order $e^{-\\beta^*}\\epsilon_n^2/\\delta^2$, improving the network-only rate $e^{-\\beta^*}d \\log n/(n\\delta^2)$ when few covariates are informative and few nodes mismatch. If correct, this gives practitioners one procedure that improves network estimation and clustering when a sparse covariate signal exists and degrades gracefully to network-only behavior when covariates are useless or misleading. Two variational algorithms, coordinate-ascent and stochastic, are developed so the method scales to networks with tens of thousands of nodes.","feed_headline":"Shrinkage priors borrow only useful covariates for clustering","feed_subtitle":"Provably matches network-only performance when covariates mislead, and beats it when a sparse set helps.","key_machinery":"The load-bearing object is the discrepancy matrix $X-ZB$ between latent network positions and a low-rank linear transformation of the covariates. The model places independent horseshoe global-local shrinkage priors, built from half-Cauchy scales, on each row $b_j$ of $B$ and on each residual $x_i-B^\\top z_i$, so the posterior can shrink irrelevant covariate rows to zero and, node by node, shrink the covariate-driven approximation toward the network-learned position only when they agree. The proof machinery is the fractional posterior analyzed through $\\alpha$-Rényi divergence, a new $\\ell_{2,1}$ prior-concentration lemma for multivariate shrinkage priors, and a Chernoff-type Gaussian concentration bound; the computation is made conjugate by the tangent transform for the Bernoulli likelihood and an inverse-Gamma mixture representation of the half-Cauchy prior. A restricted eigenvalue condition on $Z$ translates deviations in $B$ into deviations in $ZB$.","core_discovery":"The paper's central claim is that simultaneous row-sparsity on the covariate loading matrix $B^*$ and on the discrepancy matrix $X^*-ZB^*$ makes covariate-assisted network modeling adaptive: the fractional posterior recovers connection probabilities at rate $\\epsilon_n^2 = M_0^2(1+s_b+s_x)d\\log(np)/n^2$, which beats the network-only rate $d\\log n/n$ whenever $s_b+s_x=o(n)$ and $\\log p=O(\\log n)$. Under an additional projection prior that bounds and centers the latent vectors—stated to be used only in the proof, not in the algorithm—the inner products $x_i^\\top x_j$ and the intercept $\\beta$ contract with an extra $e^{-\\beta^*}$ factor, and $K$-means on the posterior estimate of $X$ misclusters at rate $O(e^{-\\beta^*}\\epsilon_n^2/\\delta^2)$, improving on the network-only rate $O(e^{-\\beta^*}d\\log n/(n\\delta^2))$. When $B^*=0$, so covariates carry no signal, the method reverts to the network-only rate, and when covariates and latent positions agree perfectly ($s_x=0$), it recovers the shared-latent-vector model. The authors further show the rate holds without requiring the covariate signal to be strong enough for exact variable selection; only the prediction error $\\|ZB-X\\|_F$ needs to be small.","pith_inferences":["The rate's dependence on $s_x$ suggests treating each mismatched node as costing one extra $d\\log(np)/n^2$ term; an implication is that the improved rate survives only when the mismatch fraction is $o(n/\\log p)$, giving a simple rule of thumb for when covariate information should be included.","Because the misclustering theorem is proved under the projected prior while the published algorithms do not enforce the projection, a direct empirical check would compare variational posterior draws against the bounded-and-centered event; if the draws frequently leave it, the formal rate should be re-derived for the unprojected family.","The same two-layer shrinkage structure could be reused for multilayer or dynamic networks where one covariate matrix informs several layers; the paper lists this as future work, and the rate analysis suggests the improvement would appear only if the shared signal is sparse in the same two-layer sense."],"forward_implications":["Link-probability estimation error drops below the network-only rate whenever $s_b+s_x=o(n)$ and $\\log p=O(\\log n)$, and this gain does not require strong-enough covariate signals for exact variable selection.","Community detection becomes consistent at cluster separation $\\delta=\\omega(e^{-\\beta^*/2}\\epsilon_n)$, which for constant $s_b,s_x$ lowers the required separation from roughly $1/\\sqrt{n}$ to roughly $1/n$ in sparse regimes.","If the covariates are pure noise, the prior drives $B$ toward zero and the method contracts at the same rate as a network-only latent-space model, so full covariate misspecification does not break the procedure.","The coordinate-ascent algorithm costs $O(n^2+p^2)$ per iteration and the stochastic variant costs $O(|E|)$ per iteration, making the approach feasible for networks with tens of thousands of nodes and high-dimensional covariates."],"supporting_citations":[{"why":"Supplies the horseshoe global-local shrinkage prior used on covariate rows and node discrepancies.","marker":"Carvalho et al. (2010)"},{"why":"Provides the fractional-posterior contraction theorem that converts the prior-mass lower bound into an $\\alpha$-Rényi rate.","marker":"Bhattacharya et al. (2019)"},{"why":"Gives the tangent transform of the Bernoulli likelihood that makes CAVI updates closed-form.","marker":"Jaakkola and Jordan (2000)"},{"why":"Represents half-Cauchy scales as inverse-Gamma mixtures, enabling conditionally conjugate variational updates.","marker":"Neville et al. (2014)"},{"why":"Motivates the restricted eigenvalue condition used as Assumption 2 to link deviations in $B$ to deviations in $ZB$.","marker":"Bickel et al. (2009)"},{"why":"Supplies the network-only latent-space baseline and the projected-prior device used in Assumption 6.","marker":"Ma et al. (2020)"},{"why":"States the network-only misclustering rate $e^{-\\beta^*}d\\log n/(n\\delta^2)$ that Theorem 3.3 is compared against.","marker":"Zhang et al. (2022a)"},{"why":"Provides the K-means perturbation bound used as Lemma A.7 in the community-detection proof.","marker":"Lei and Rinaldo (2015)"}],"fun_headline_variants":["Adaptive network modeling: borrow only helpful covariates","Shrinkage priors pick useful covariates, ignore the rest","Robust latent space model adapts to misleading covariates","Network clustering that provably ignores bad covariates","Selective covariate borrowing for robust network clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formal latent-vector and clustering guarantees hold for a projected prior that bounds and centers the latent vectors, but the published algorithms do not apply that projection; if the projection materially changes posterior behavior, the improved rates may not hold for the method as actually run.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive network modeling: borrow only helpful covariates","Shrinkage priors pick useful covariates, ignore the rest","Robust latent space model adapts to misleading covariates","Network clustering that provably ignores bad covariates","Selective covariate borrowing for robust network clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1958,"prompt_tokens":1029,"completion_tokens":929,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":855}},"tokens_in":645,"tokens_out":929,"duration_ms":9207,"temperature":1.0,"reasoning_tokens":855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:38:03.103159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate networks from the model with true latent vectors that violate the boundedness-and-centering event, for example by letting their norms grow with $n$ or giving them a nonzero row mean, then fit the published CAVI/SVI algorithms and measure misclustering as $n$ grows; if the error does not track $e^{-\\beta^*}(1+s_b+s_x)d\\log(np)/(n^2\\delta^2)$ but instead tracks the network-only $e^{-\\beta^*}d\\log n/(n\\delta^2)$, the projection assumption is load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fractional-posterior contraction theorem that converts the prior-mass lower bound into an $\\alpha$-Rényi rate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the tangent transform of the Bernoulli likelihood that makes CAVI updates closed-form."},{"cited_title":"E., Ormerod, J","cited_arxiv_id":null,"evidence_quote":"Represents half-Cauchy scales as inverse-Gamma mixtures, enabling conditionally conjugate variational updates."},{"cited_title":"J., Ritov, Y., and Tsybakov, A","cited_arxiv_id":null,"evidence_quote":"Motivates the restricted eigenvalue condition used as Assumption 2 to link deviations in $B$ to deviations in $ZB$."}],"review_version":1}