{"id":"368cdb0d-6545-427f-8320-efb62b885182","arxiv_id":"2508.16556","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A spherical latent space model for undirected binary networks is proposed, using cosine similarity between unit-vector positions, with Bayesian MCMC inference and comparisons on Florentine marriage and monk network data.","lead":"This paper introduces a latent space model that places social network actors on a sphere, so the probability of a tie depends on how close two actors are on the sphere. The authors test it on two classic small networks and report better predictive scores than flat-space models, but their model-fit table tells a more mixed story.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical case for S2 rests on an undocumented out-of-sample protocol with hyperparameters tuned on predictive criteria, while the paper's own WAIC contradicts improved fit; a nested CV check is needed to determine whether the predictive advantage survives.","rationale":"The reader's verdict identifies the most load-bearing weakness: Table 4 is presented without a documented evaluation protocol and the hyperparameters are tuned on predictive criteria, making the headline predictive claim unfalsifiable from the paper alone. I agree that this warrants rejection of the central empirical claim. I additionally note that Table 2 contradicts the 'improved model fit' assertion and contains a suspicious internal inconsistency (ML log-likelihood below posterior-mean log-likelihood for S2), further undermining the reliability of the reported numerical comparisons. The proposed model and inference machinery are reasonable and clearly presented, so the issue is evidentiary rather than mathematical; the verdict should remain REJECT until the protocol is specified and the comparison is rerun under a valid nested evaluation.","tokens_in":20070,"tokens_out":6205,"duration_ms":80092,"concrete_test":"Inspect the repository at https://github.com/cnosa/LatentSpaces_Network_Manifold and identify the exact out-of-sample protocol behind Table 4. Then rerun the Florentine comparison with nested 5-fold cross-validation: tune the hyperparameters (σz, σα, ρ, µβ, σβ) on training folds only, and evaluate S2, R3, and a Euclidean inner-product model on held-out edges, reporting mean and standard deviation of AUC and F1 across folds. If S2 no longer dominates, or the margins fall within fold-to-fold noise, the claimed predictive advantage is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the spherical model yields improved fit and prediction is not supported by the evidence in the manuscript. Section 5 reports Table 4 as 'out-of-sample AUC, accuracy, and F1' but never states the train/test split, number of folds, edge-sampling scheme, or whether the same folds were used across models. The only protocol detail is that hyperparameters were 'selected via sensitivity analysis to optimize information and predictive criteria (supplementary material)', so if the predictive criteria in Table 4 are the same held-out criteria, the S2 advantage may be a tuning artifact. Table 2 independently undercuts the 'improved model fit' claim: S2's WAIC is 199.367 versus R3's 161.099. Table 2 also contains an internal inconsistency: for S2, the reported ML log-likelihood is -31.852, far below the posterior-mean log-likelihood of -13.469, which is impossible if the reported MLE is a maximizer; this indicates that the numerical comparisons are unreliable. Finally, the design confounds geometry with link function: S2 uses a symmetric inner-product link while the Euclidean baselines use distance, so even a genuine predictive gain would not isolate the value of spherical geometry.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spherical latent space model for undirected binary social networks, where nodes are embedded on S^{d-1} and tie probabilities depend on the inner product of latent positions. Bayesian inference is carried out with Metropolis–Hastings, Hamiltonian Monte Carlo, and geodesic/spherical HMC, with an application to the Florentine families network and a monks network in the supplement. The authors claim that the spherical geometry yields improved model fit, predictive performance, and interpretability relative to standard Euclidean latent distance models.","tokens_in":20421,"tokens_out":4786,"duration_ms":59322,"significance":"If the empirical claims were substantiated, this would be a useful contribution: the spherical constraint removes translation and scale indeterminacies, gives a compact bounded latent space, and connects latent space modeling to directional statistics. The likelihood and gradient derivations are standard and appear correct, and the authors include reproducible code. However, the central evidence in the manuscript is not currently convincing: the paper's own model comparison table does not support the 'improved model fit' claim, the predictive evaluation protocol is undocumented, and the model comparison confounds geometry with the choice of link function.","major_comments":[{"comment":"The abstract and Section 5 claim 'improved model fit' for the spherical model, but Table 2 shows S2 WAIC = 199.367, which is worse than R2 (170.072) and R3 (161.099), and S1 WAIC = 384.106 is the worst overall. The statement that 'spherical models achieve log-likelihood values close to their Euclidean counterparts' does not address the WAIC ranking, and S2 has fewer effective parameters than R3 (n(d-1)+2 vs. nd+1), so the WAIC gap is not explained by complexity. This directly contradicts a central claim of the paper.","section":"Section 5, Table 2"},{"comment":"Table 2 contains internally inconsistent log-likelihood values. For S2, the ML log-likelihood is -31.852, but the MAP log-likelihood is -16.570 and the posterior-mean log-likelihood is -13.469, both exceeding the maximum of the log-likelihood. For R1, the mean log-likelihood (-43.105) also exceeds the ML log-likelihood (-47.632). Since the MLE maximizes the log-likelihood by definition, these numbers indicate an error in the optimization, MCMC post-processing, or table construction. The numerical comparisons involving S2 are therefore unreliable and must be corrected before the empirical claims can be evaluated.","section":"Section 5, Table 2"},{"comment":"The out-of-sample predictive evaluation behind Table 4 is not described. The text states that AUC, accuracy, and F1 were assessed 'out-of-sample' and that hyperparameters were 'selected via sensitivity analysis to optimize information and predictive criteria (supplementary material)', but it does not state the train/test split, the number of folds, the edge-sampling scheme, or whether the same folds were used for all models. If the predictive criteria used for tuning are the same as those reported in Table 4, the S2 advantage may be a tuning artifact. The full protocol must be reported, or the strong claim that 'S2 outperforms all others' should be withdrawn.","section":"Section 5, Table 4"},{"comment":"The comparison confounds geometry with link function. The spherical model uses the symmetric inner-product link η_{ij} = α + β⟨z_i, z_j⟩, while the Euclidean baselines use the distance link η_{ij} = α - ||z_i - z_j||. Any observed difference in fit or prediction could be due to the change of link function rather than the spherical geometry. To support the claim that spherical geometry itself is beneficial, the paper should include a Euclidean inner-product/projection baseline (for example, the projection model of Hoff et al. 2002) or otherwise isolate the geometric effect.","section":"Section 3, model specification"}],"minor_comments":[{"comment":"The column headers of Table 4 read 'R1  R2  R2  S1  S2'; the third column should presumably be 'R3'.","section":"Section 5, Table 4"},{"comment":"Typo: 'see Section 3.1 dor details' should be 'for details'.","section":"Section 3.1"},{"comment":"The text 'andacorrelatedGaussianprior' is missing spaces; it should read 'and a correlated Gaussian prior'.","section":"Section 4.2"},{"comment":"Line 14 of Algorithm 4 uses 'p(m)' which is undefined; it should likely be 'p^{(b)}' or 'p*'.","section":"Section 4.2.1, Algorithm 4"},{"comment":"The caption says 'Latent space representations for R1 and S1 models', but the figure contains panels for R1, S1, R2, and S2. The caption should list all four panels.","section":"Section 5, Figure 5"},{"comment":"The sensitivity analysis mentioned in the text is only described as 'supplementary material'; the main text should at least summarize the ranges of hyperparameters considered and the criterion used.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The model and algorithms are coherent, and the manuscript includes code, but the central empirical claims are not supported by the evidence as presented. The inconsistencies in Table 2 and the missing predictive protocol are fixable, so I do not recommend rejection; however, the authors need to either correct the numerical results and provide a full out-of-sample protocol, or substantially weaken the claims about improved fit and prediction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The model is a straightforward but clean extension of Hoff et al. (2002): put latent positions on the unit sphere and use a symmetric inner-product link instead of Euclidean distance. The authors are honest that Yu (2020) did spherical latent factors for roll-call data, so the novelty is the application to undirected social networks. The math is careful—gradients, identifiability up to rotation, and the GHMC details are all standard and correct. The code is available, which is a plus.\n\nThe soft spots are empirical and load-bearing. The abstract claims improved fit, but Table 2 shows WAIC of 199.4 for S2 versus 161.1 for R3. That is a direct contradiction. The predictive results in Table 4 favor S2, but there is no description of the train-test split, number of folds, or edge-sampling scheme. Hyperparameters were tuned via sensitivity analysis to optimize predictive criteria on the same data, so the gain could be a tuning artifact. On top of that, Table 2 has an internal inconsistency: for S2, the ML log-likelihood is -31.852 and the posterior-mean log-likelihood is -13.469. That is impossible if the ML estimate actually maximizes the likelihood; either the optimization failed or the posterior mean is being evaluated incorrectly. Finally, the comparison confounds geometry with link function—S2 uses inner product, Euclidean baselines use distance—so even a clean advantage wouldn't isolate the value of spherical geometry.\n\nThe paper is well written and the model is reasonable. The problem is that the central empirical claim is not supported by the evidence as presented. I would send it to peer review because it deserves a serious referee who can demand a proper evaluation protocol. With a corrected numerical table and a nested cross-validation comparison that holds the link function fixed, it could be a useful applied contribution. As is, I would not cite it or use it as evidence for spherical geometry.","headline":"A clean spherical latent space model with sound math, but the empirical superiority claims are contradicted by the paper's own WAIC table and an undocumented predictive protocol.","tokens_in":20872,"tokens_out":3074,"would_cite":false,"duration_ms":32961,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that placing a social network's hidden 'social space' on a unit sphere, so that tie probability rises with cosine similarity between latent positions, improves link prediction and community detection over the standard Eucl","keywords":["latent space models","social network analysis","spherical geometry","Bayesian inference","Markov chain Monte Carlo","link prediction","community detection","von Mises–Fisher"],"falsifier":"Run a pre-registered ten-fold cross-validation on the Florentine families and monks networks, using identical priors and identical hyperparameter-selection rules for all models, and report the S2-versus-R3 difference in AUC and F1; the paper's Table 4 does not state the data split, so this check would settle whether S2's reported 0.994 AUC is a real out-of-sample gain or a tuning artifact.","tokens_in":19971,"feed_emoji":"🌐","tokens_out":11747,"duration_ms":121831,"temperature":0.7,"pith_summary":"The paper aims to show that the hidden geometry behind a social network can usefully be spherical rather than flat. It proposes a latent space model in which each actor has a position on the unit sphere S^{d-1}, and the log-odds of a tie between two actors is α plus β times the cosine of the angle between their positions, so angular alignment, not Euclidean distance, drives connection probability. The authors develop Bayesian inference with uniform priors on the sphere, von Mises–Fisher proposals, and geodesic Hamiltonian Monte Carlo, and they demonstrate the model on the Florentine families marriage network. They report that the two-dimensional spherical model S2 gives the best out-of-sample AUC, accuracy, and F1 among all Euclidean and spherical alternatives, while producing tighter, more interpretable embeddings and higher-modularity community partitions. If that holds, analysts gain a bounded, low-parameter geometry that naturally represents transitive and cyclical structure without the degeneracy risks of unbounded Euclidean embeddings.","feed_headline":"Spheres beat flat space for network link prediction","feed_subtitle":"A spherical latent-space model claims better out-of-sample accuracy and clearer communities on classic social networks.","key_machinery":"The load-bearing object is the spherical latent space model S^{d-1}: each actor i carries a unit vector z_i on the (d−1)-sphere, and the linear predictor is η_ij = α + β z_i^T z_j. This inner-product link is the mechanism that converts angular similarity into tie probability; because z_i^T z_j = cos γ_ij, it gives a bounded, rotation-invariant geometry with every position having only d−1 free parameters. The companion machinery is the Bayesian sampler: uniform priors on the sphere, von Mises–Fisher proposals for single-position Metropolis updates, and geodesic Hamiltonian Monte Carlo for manifold-aware exploration.","core_discovery":"The central claim is that switching the latent space of a social network model from Euclidean space to the unit sphere changes the model's bias in a beneficial way. The authors specify η_ij = α + β⟨z_i, z_j⟩ with ∥z_i∥ = ∥z_j∥ = 1, so a tie is more probable when two actors' latent directions align; because ⟨z_i, z_j⟩ = cos(γ_ij), this is equivalent to a decreasing function of geodesic distance, but the cosine form avoids arccos instabilities. The sphere fixes origin and scale, leaving only global rotations as a non-identifiability, and it bounds all distances, which the authors argue prevents the extreme tie probabilities and norm drift that can plague Euclidean latent distance models. In th","pith_inferences":["A natural but untested consequence is that the spherical model's advantage is strongest on networks with cyclic or periodic tie patterns—dominance hierarchies, citation cycles, seasonal interaction patterns—where circular geometry matches the topology; the paper motivates this but does not analyze such data.","The paper's predictive table and its in-sample WAIC table point in opposite directions (WAIC favors R3; AUC/F1 favor S2), so a careful reader should read the abstract's 'improved model fit' as shorthand for out-of-sample prediction and visualization, not within-sample likelihood.","If the spherical advantage holds, the obvious next experiment is a head-to-head against hyperbolic latent spaces, since both geometries bound distances and hyperbolic space is tailored to hierarchical trees; the paper's closing discussion invites this but does not run it.","Because the inner-product link is symmetric, the model cannot yet represent directed relations; a testable extension is to replace the inner product with an asymmetric kernel or weighted bilinear form and ask whether the spherical constraint still beats Euclidean baselines."],"forward_implications":["Spherical latent positions eliminate unbounded-distance degeneracy, so optimization and MCMC avoid the norm drift that can destabilize Euclidean latent-space fits.","On the Florentine families network, the S2 model reports the best out-of-sample AUC (0.994), accuracy (0.952), and F1 (0.857) among R1, R2, R3, S1, and S2, suggesting the geometry itself improves link prediction on dense, alliance-driven networks.","Spherical embeddings produce tighter posterior configurations (S2 average posterior variance 0.154 vs 3.157 for R3), giving sharper, more compact latent-space visualizations.","Community detection on the spherical latent spaces achieves higher modularity (S1: 0.391) than on Euclidean spaces (best Euclidean R2: 0.351), so the geometry helps expose block structure.","The same spherical construction extends to mixed-effect covariates, weighted inner products, finite mixtures, Dirichlet-process clustering, multilayer networks, and dynamic networks, all spelled out in the paper's Section 3.2."],"supporting_citations":[{"why":"Defines the Euclidean latent distance model and projection model that the paper extends, and supplies the benchmark datasets for the illustrations.","marker":"Hoff et al. (2002)"},{"why":"Introduces spherical latent factor models for binary and ordinal data and the spherical Hamiltonian Monte Carlo sampler that the paper adapts to network data.","marker":"(Yu, 2020)"},{"why":"Provides the Florentine families marriage network and the historical narrative used to interpret the inferred central role of the Medici.","marker":"Padgett and Ansell (1993)"},{"why":"Supplies the simulation algorithm for von Mises–Fisher proposals used to update spherical latent positions in the Metropolis sampler.","marker":"Wood (1994)"},{"why":"Establishes the Riemannian manifold Hamiltonian Monte Carlo framework that the geodesic Hamiltonian Monte Carlo algorithm builds on.","marker":"(Girolami and Calderhead, 2011)"},{"why":"The spectral clustering method used to detect communities in the inferred latent spaces and compute modularity comparisons.","marker":"(Ng et al., 2002)"},{"why":"The greedy modularity maximization used in posterior predictive checks for observed and simulated networks.","marker":"(Clauset et al., 2004)"}],"fun_headline_variants":["Spherical latent space boosts social network fit","On the sphere: better ties, clearer communities","Curved latent space beats Euclidean for networks","Sphere model curbs degeneracy, improves ties"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper's empirical case for the spherical model rests on the unstated evaluation protocol behind Table 4 and on prior hyperparameters tuned to optimize predictive criteria, so if the reported AUC and F1 for S2 are not from a genuinely out-of-sample split with fairly tuned hyperparameters, the predictive advantage would be an artifact—and the paper's own WAIC table, which favors Euclidean R3, already shows that the advantage is not one of in-sample fit.","fun_headline_variants_meta":{"raw":{"variants":["Spherical latent space boosts social network fit","On the sphere: better ties, clearer communities","Curved latent space beats Euclidean for networks","Sphere model curbs degeneracy, improves ties"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1004,"prompt_tokens":621,"completion_tokens":383,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":365,"completion_tokens_details":{"reasoning_tokens":326}},"tokens_in":365,"tokens_out":383,"duration_ms":4910,"temperature":1.0,"reasoning_tokens":326,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:13:21.000844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a pre-registered ten-fold cross-validation on the Florentine families and monks networks, using identical priors and identical hyperparameter-selection rules for all models, and report the S2-versus-R3 difference in AUC and F1; the paper's Table 4 does not state the data split, so this check would settle whether S2's reported 0.994 AUC is a real out-of-sample gain or a tuning artifact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Florentine families marriage network and the historical narrative used to interpret the inferred central role of the Medici."}],"review_version":1}