{"id":"5af9c00d-14c6-47cb-80ec-f25b474d584a","arxiv_id":"2505.22289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"NetMA averages latent space models with different dimensions using K-fold edge cross-validation, and is claimed to be asymptotically optimal for link prediction while outperforming true-model fitting when the true dimension is large.","lead":"NetMA averages latent space models with different dimensions for network link prediction, choosing weights by K-fold edge cross-validation. If correct, it gives a principled alternative to selecting one latent dimension, and simulations claim it can beat the true-dimension oracle when the true dimension is large.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The f1 correction in K-fold edge CV omits the 1/p factor from the Ψ1/Ψ2 split, so the CV criterion's pseudo-true limit is pP* rather than P*; Theorem 1's equivalence between CV and L(w) fails as stated.","rationale":"The paper's central theoretical claim is that K-fold edge CV weights asymptotically minimize the prediction loss L(w). This requires the CV criterion to be asymptotically equivalent to L(w), which in turn requires the hold-out predictions eP^[k] and the final predictions bP to have the same pseudo-true limits. The displayed formulas violate this: bP corrects for the initial Ψ1/Ψ2 split by dividing by p, while eP^[k] corrects only for the K-fold split, giving limit pP* instead of P*. The assertion in Section 2.3 that the two have the same limiting values is thus incorrect, and the omitted supplementary proof cannot fix this without changing the method. This is more concrete than the reader's high-level concern about Assumptions 1-2: even granting all the growth conditions, the criterion being minimized is not the right one. The multi-layer analogue suffers the same problem. I would not reject the underlying idea: with the corrected factor f1(x) = K/[p(K-1)(1+e^{-x})], the procedure is coherent and the existing proof strategy might go through. But the version as written is internally inconsistent, and the reported simulation advantage over the oracle could be an artifact of the biased CV. A targeted re-derivation and one simulation rerun would settle this. I also credit the paper for a clear problem formulation and a sensible overall averaging framework; the issue is in the implementation detail of the CV correction, not in the conceptual approach.","tokens_in":26032,"tokens_out":17325,"duration_ms":181423,"concrete_test":"Derive the pseudo-true limit of eP^[k]_(m) in a homogeneous network with all P_ij = 0.5, p = 0.7, K = 10, and one candidate model. The displayed f1 gives limit p·0.5 = 0.35, not 0.5, confirming the missing 1/p factor. Then rerun the single-layer simulation Case 1 (N = 200, d0 = 10, M = 12, K = 10, p = 0.7) with the corrected factor f1(x) = K/[p(K-1)(1+e^{-x})] and compare NetMA's relative risk and selected weights; material changes would indicate the reported advantage over the oracle is an artifact of the biased CV.","verdict_should_be":"REJECT","load_bearing_attack":"In Section 2.2, the final predictor is bP_(m) = bP_0_(m)/p with p = |Ψ1|/|Ψ|, so its pseudo-true limit is P*_(m). In Step 2(b), the hold-out predictions are eP^[k]_(m) = f1(θ^[-k]) with f1(x) = K/{(1+e^{-x})(K-1)}. The fitting set for fold k has effective sampling probability q = p(K-1)/K, so the raw fitted probabilities σ(θ^[-k]) converge to qP*_(m). Hence eP^[k]_(m) converges to [K/(K-1)]·qP*_(m) = pP*_(m), not P*_(m). The paper's assertion in Section 2.3 that bP_(m) and eP^[-k]_(m) 'have the same limiting values P*_(m),ij' is therefore false whenever p < 1. The K-fold CV criterion CV(w) is then (1/|Ψ1|)Σ||A[k] - pP*(w)||^2, whose minimizer does not equal the minimizer of L(w) on Ψ2; the simplex constraint makes the p-scaling non-invariant. Consequently Theorem 1's claim that L(w_hat)/inf_w L(w) -> 1 is not supported by the stated algorithm, since the proof would have to rely on eP and bP having the same limit, which they do not. The same omitted 1/p factor appears in the multi-layer CV definition in Section 3.2, undermining Theorems 4-5 as well. This is an internal inconsistency visible directly from the displayed formulas, not merely a missing primitive condition.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a frequentist model averaging method (NetMA) for link prediction in single-layer and multi-layer latent space models with varying latent dimensions. The weights are selected by minimizing a K-fold edge cross-validation criterion, and the final prediction averages the debiased probability estimates from all candidate models. The paper claims three theoretical results: asymptotic optimality of the CV-selected weights relative to the infeasible best squared-error loss when all candidate models are misspecified, consistency of the weights toward correct models when such models are included, and a convergence rate of the CV weights to the infeasible optimal weight vector. These claims are stated for both single-layer and multi-layer networks. The paper also reports extensive simulations and two real-data link prediction applications.","tokens_in":26614,"tokens_out":12280,"duration_ms":142522,"significance":"If the theoretical results were correct, the paper would fill a genuine gap: it extends frequentist optimal model averaging from independent data to network data, where the dependence structure and one-realization setting make edge-based cross-validation natural. The computational proposal is simple (quadratic programming), and the simulations cover a useful range of settings, including covariates and egocentric missingness, with a transparent evaluation design. The empirical evaluation is a strength of the manuscript. However, the key theoretical claims are currently not supported because of an internal inconsistency in the CV correction factor (see Major Comment 1), and the proofs are not available in the posted version. The paper's significance is therefore conditional on repair of that factor and on provision of the supporting proofs.","major_comments":[{"comment":"The definition f1(x) = K/{(1+e^{-x})(K-1)} omits the factor 1/p. In fold k the fitting set is Ψ1 \\ Gk, whose effective sampling proportion is q = p(K-1)/K; after fitting on A(Ψ1\\Gk), the raw fitted probabilities σ(θ^[-k]) converge to qP*(m) under the paper's own debiasing logic, so eP[k](m) converges to [K/(K-1)] qP*(m) = pP*(m), not P*(m). The statement in Section 2.3 that bP(m),ij and eP[-k](m),ij have the same limiting values is therefore false for p < 1. Consequently CV(w) asymptotically minimizes a criterion proportional to ||P - pP*(w)||^2 on the held-out groups rather than L(w), and since the unit-simplex constraint is not invariant to scaling the candidate probabilities, the minimizer of CV(w) need not coincide with the minimizer of L(w). This invalidates the proof route of Theorem 1 and, by the same argument, Theorems 2 and 3; the identical missing factor in Section 3.2 undermines Theorems 4-6. The fix is to replace f1(x) by K/{p(1+e^{-x})(K-1)} and to re-derive the proofs with that correction.","section":"Section 2.2, Step 2(b); Section 3.2, Step 2(b)"},{"comment":"Theorems 1-6 and the verification of Assumptions 1-3 are deferred to a supplementary file that is not included in the arXiv v1. Since the central contribution of the paper is these asymptotic results, the current version cannot be independently verified. Please include the proofs in the revised version or provide the supplement; this is a prerequisite for acceptance.","section":"Supplementary Material"},{"comment":"The assumptions are stated at a high level (existence of pseudo-true limits and growth-rate conditions on ξ* and ξ*†) without primitive conditions on the latent space model or the sampling scheme. The growth condition in Assumption 2 explicitly rules out the case where any candidate model is correct, so Theorem 1 does not cover the small-N large-d0 regime in which the simulations show the largest gains over the oracle. If the supplementary material verifies these assumptions from primitive conditions, please state that explicitly in the main text; otherwise the scope of the theoretical results should be narrowed accordingly.","section":"Section 2.3, Assumptions 1-2; Section 3.3, Assumptions 7-9"}],"minor_comments":[{"comment":"The definition of the squared risk R†(w) is missing the square: it should read R†(w) = Σ_t Σ_{(i,j)∈Ψ1} E( bP(t)ij(w) - P(t)ij)^2.","section":"Section 3.3"},{"comment":"The text refers to 'Table 6' for the ResearchGate link prediction results, but the table is labeled Table 4; please correct the cross-reference.","section":"Section 5.1"},{"comment":"There are minor typographical issues: 'matrics' should be 'matrices', and the reference to Roozbahani et al. contains 'reseachgate', which should be 'ResearchGate'.","section":"Throughout"},{"comment":"The term 'oracle' in the simulations denotes the estimator using the true latent dimension, not the infeasible optimal weight vector w0; the phrase 'outperforms the oracle' should be qualified to avoid confusion with the oracle averaging predictor used in the theory.","section":"Section 4.1"},{"comment":"The notation for the bias correction is applied inconsistently: the final estimator bP(m) divides by p, while the hold-out estimator eP[k](m) uses f1 without the p correction; even after the correction in Major Comment 1, please state explicitly that the same sampling fraction is used in both places.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The missing 1/p factor in f1 is a concrete, fixable error, but it currently undermines the main theoretical theorems. The empirical sections appear carefully run, and the method is potentially useful, so I do not recommend rejection. Please ensure the supplementary proofs are available to the referees in the next round and that the assumptions are verified from primitive conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"NetMA is a new and useful idea: K-fold edge cross-validation to average latent space models of different dimensions for link prediction, with a clean split between weight selection and evaluation. The simulation results are plausible and the empirical work is thorough. The writing is clear.\n\nThe soft spot is load-bearing. The hold-out prediction in Step 2(b) uses f1(x)=K/{(1+e^{-x})(K-1)}, but the training set for each fold is only a fraction p(K-1)/K of all edges. The fitted probabilities on that set therefore converge to qP* with q=p(K-1)/K, and after the f1 correction they converge to pP*, not P*. The final predictor bP_(m)=bP0_(m)/p does converge to P* (under Assumption 1), so the target loss L(w) is based on P*, while the CV criterion is based on pP*. Minimizing ||P - pP*(w)||^2 is not the same as minimizing ||P - P*(w)||^2 on the simplex, so Theorem 1's claim L(bw)/inf L(w) -> 1 does not follow from the stated algorithm. The same missing 1/p factor appears in the multi-layer CV criterion, undermining Theorems 4-5 as well. This is visible directly from the displayed formulas, and it is not a missing primitive condition.\n\nTwo smaller issues: the proofs of Theorems 1-6 are deferred to a supplementary file not included in the arXiv v1, and Assumptions 1-13 are stated at a high level with no primitive checks. The simulation figures in the main text have no error bars, though the empirical tables do.\n\nNone of this kills the underlying idea. The method is sensible and the sims are consistent with its motivation; a corrected correction factor and a full proof file could put the theorems on solid ground. But as it stands, the central optimality claim is unsupported.\n\nI'd send this to a serious referee—the novelty justifies the time—but tell the referee to hammer on the f1 scaling before looking at anything else.","headline":"NetMA is a genuinely new model-averaging proposal for latent space link prediction, but the displayed CV formulas have a scaling mismatch (missing 1/p) that undermines the asymptotic optimality theorem as stated.","tokens_in":26941,"tokens_out":7121,"would_cite":false,"duration_ms":78372,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62H12","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Edge cross-validation weights make averaged latent-space link prediction asymptotically optimal, and often better than the true-dimension oracle.","keywords":["network model averaging","latent space model","link prediction","edge cross-validation","asymptotic optimality","multi-layer networks","model averaging weights","K-fold cross-validation"],"falsifier":"Fit the candidate family to a network generated from the model with known P for increasing N, with the true dimension included among the candidates, and compute L(ŵ)/inf_w L(w) together with the weight assigned to the correct dimension: Theorem 1 predicts the ratio tends to 1 and Theorem 2 predicts all weight lands on the correct models, so a simulation where the ratio stays above 1 or the weight fails to concentrate would refute the claims. A sharper probe is a dense network where the minimal averaged loss does not grow with N, a setting the assumptions exclude; if the ratio still converges to 1 there, the conditions are stronger than needed, and if it does not, the stated boundary of the theory is confirmed.","tokens_in":25807,"feed_emoji":"🔗","tokens_out":9129,"duration_ms":86046,"temperature":0.7,"pith_summary":"Networks with few nodes and many latent dimensions are hard to model: the true dimension is unknown, and picking one dimension risks choosing badly. This paper proposes NetMA, which fits latent space models at several candidate dimensions and averages their link predictions using weights chosen by K-fold edge cross-validation, splitting pairs of nodes rather than nodes themselves so the network structure survives. The central theoretical claim is that these edge-CV weights are asymptotically optimal: the averaged predictor's squared loss converges to the lowest loss achievable by any weighting of the candidates, even when every candidate dimension is misspecified. When the candidate set includes correctly specified models, the weights concentrate on those models. The practical payoff is that averaging can beat the oracle procedure that knows the true dimension, exactly in the small-network, large-dimension regime where single-model estimation is unstable.","feed_headline":"Average across latent dimensions to beat the oracle at link prediction","feed_subtitle":"New K-fold edge-cross-validation weights make averaged network fits provably no worse than the best candidate model.","key_machinery":"The load-bearing object is the K-fold edge cross-validation criterion. The method splits the set of node pairs symmetrically into a fitting part and a prediction part, fits each candidate model on the fitting part with a bias correction (dividing fitted probabilities by the inclusion probability p), then splits the fitting pairs into K folds. Each candidate is refit leaving one fold out, and the held-out fold is predicted through the adjusted link function f1(x) = K/{(1+$e^{{-x}}$)(K-1)} that compensates for the missing folds. The criterion CV(w) = (1/|Ψ1|) Σ_k ||A[k] − P̃[k](w)||$_F^{2}$ is a quadratic function of the weight vector w, so selecting weights reduces to a quadratic program over the unit simplex. All asymptotic optimality statements hinge on this criterion tracking the unobservable squared loss L(w) on the held-out pairs; the debiasing factors and the pair-splitting design are what make edge CV, rather than node CV, the appropriate loss estimator for network data.","core_discovery":"For an undirected network with connection probabilities logit(P_ij) = α_i + α_j + z_i^T z_j, averaging predictions across latent space models of dimensions 1 through M with weights from K-fold edge cross-validation is asymptotically no worse than the best possible weighted average of those models. Theorem 1 states that L(ŵ)/inf_w L(w) → 1 in probability under the stated assumptions, where L is the squared prediction loss on held-out node pairs; Theorem 2 states that when some candidate dimensions are at least the true dimension, the total weight on correct models tends to 1. The same pair of results holds for multi-layer networks (Theorems 4 and 5), and the empirical weight vector converges to the infeasible optimal weight vector at rate O_p($ξ^{{1/2}}$|Ψ1|^{-1/2+κ}) (Theorems 3 and 6). Simulations show the method outperforming the oracle when the true latent dimension is large relative to network size.","pith_inferences":["The oracle-beating phenomenon points to a regularization mechanism: averaging over misspecified low-dimensional fits trades a little bias for a large cut in estimation variance, and the paper does not formalize this bias-variance decomposition.","Because NetMA only consumes fitted edge probabilities, it should transfer to block models, universal singular value thresholding, or other edge-probability estimators, with the bias-correction factors adjusted to each estimator's link function.","A natural test is to apply the same edge-CV weighting to directed or dynamic latent space models, where estimation instability is even more pronounced, to see whether averaging helps there as it does here."],"forward_implications":["Practitioners can include a range of candidate latent dimensions and let the edge-CV weights decide, removing the need to know or select the true dimension for link prediction.","When the true dimension is among the candidates, the weights concentrate on it asymptotically, so averaging hedges against picking the wrong dimension in finite samples without sacrificing consistency.","The multi-layer guarantee extends the method to networks sharing nodes across layers, with a common degree parameter and layer-specific latent positions.","Averaging can beat the true-dimension oracle in small-N, large-d0 settings, making NetMA a credible default for link prediction on modest-sized networks.","Because the criterion only uses fitted edge probabilities, the weighting scheme can be applied to other network estimators for undirected unweighted networks, not only latent space models."],"supporting_citations":[{"why":"Supplies the latent space model with logit link that all candidate models in the paper are built on.","marker":"Hoff et al. (2002)"},{"why":"Provides the edge-sampling splitting scheme that the K-fold edge cross-validation adapts to networks.","marker":"Li et al. (2020)"},{"why":"Supplies the projected gradient descent algorithm used to fit each candidate latent space model.","marker":"Ma et al. (2020)"},{"why":"Provides the multi-layer latent space model estimation that Algorithm 1 extends.","marker":"Zhang, Xue & Zhu (2020)"},{"why":"The K-fold cross-validation model averaging framework whose optimality theory NetMA transfers to network data.","marker":"Zhang & Liu (2023)"},{"why":"Discusses edge cross-validation for model selection and serves as the comparison baseline in simulations.","marker":"Gao & Ma (2020)"},{"why":"The template for the growth-rate assumption on the minimal loss used in Assumption 2.","marker":"Ando & Li (2017)"},{"why":"The template for the singular-value condition on the prediction design matrix used in Assumption 4.","marker":"Liao & Zou (2020)"}],"fun_headline_variants":["Averaging latent dimensions with edge CV provably beats the oracle","K-fold edge CV weights make latent model averaging optimal for link prediction","Edge CV weight averaging beats every single latent dimension model","NetMA: edge CV averaging of latent models achieves optimal link prediction","Provably optimal link prediction via edge-CV weighted latent model averaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central guarantee rests on growth-rate conditions: the best possible squared prediction error from averaging must grow faster than the number of candidate models times the number of node pairs, and each fitted probability must converge to a fixed limit at that speed — conditions that are assumed rather than proven from the model, and that automatically fail when one candidate model is exactly correct.","fun_headline_variants_meta":{"raw":{"variants":["Averaging latent dimensions with edge CV provably beats the oracle","K-fold edge CV weights make latent model averaging optimal for link prediction","Edge CV weight averaging beats every single latent dimension model","NetMA: edge CV averaging of latent models achieves optimal link prediction","Provably optimal link prediction via edge-CV weighted latent model averaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00034,"raw_usage":{"total_tokens":1885,"prompt_tokens":963,"completion_tokens":922,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":834}},"tokens_in":579,"tokens_out":922,"duration_ms":7868,"temperature":1.0,"reasoning_tokens":834,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:11:27.230647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the candidate family to a network generated from the model with known P for increasing N, with the true dimension included among the candidates, and compute L(ŵ)/inf_w L(w) together with the weight assigned to the correct dimension: Theorem 1 predicts the ratio tends to 1 and Theorem 2 predicts all weight lands on the correct models, so a simulation where the ratio stays above 1 or the weight fails to concentrate would refute the claims. A sharper probe is a dense network where the minimal averaged loss does not grow with N, a setting the assumptions exclude; if the ratio still converges to 1 there, the conditions are stronger than needed, and if it does not, the stated boundary of the theory is confirmed.","supporting_citations":[],"review_version":1}