Spherical latent space models for social network analysis

Carlos Nosa; Juan Sosa

REVIEW 4 major objections 6 minor 1 cited by

Spherical latent space models for social network analysis

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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

desk verdict 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. read the letter →

arxiv 2508.16556 v1 pith:4LVWFKVC submitted 2025-08-22 stat.ME stat.CO

Juan Sosa , Carlos Nosa This is my paper · ORCID

classification stat.MEstat.CO

keywords latentspacemodelssocialnetworkanalysissphericalgeometryBayesianinferenceMarkovchainMonteCarlolinkpredictioncommunitydetectionvonMises–Fisher

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

The 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.

Referee Report

4 major / 6 minor

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.

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 (4)

[Section 5, Table 2] 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 5, Table 2] 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 5, Table 4] 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 3, model specification] 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.

minor comments (6)

[Section 5, Table 4] The column headers of Table 4 read 'R1 R2 R2 S1 S2'; the third column should presumably be 'R3'.
[Section 3.1] Typo: 'see Section 3.1 dor details' should be 'for details'.
[Section 4.2] The text 'andacorrelatedGaussianprior' is missing spaces; it should read 'and a correlated Gaussian prior'.
[Section 4.2.1, Algorithm 4] Line 14 of Algorithm 4 uses 'p(m)' which is undefined; it should likely be 'p^{(b)}' or 'p*'.
[Section 5, Figure 5] 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 5] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the spherical model is a self-contained empirical proposal; the predictive evaluation protocol is under-specified but not shown to be circular.

full rationale

The paper's core contribution is an empirical model specification: latent positions constrained to a sphere, with a well-defined likelihood, prior, and MCMC inference. There is no derivation that reduces a predicted quantity to an input by construction. The spherical linear predictor eta = alpha + beta <z_i, z_j> is defined independently of the Euclidean distance baseline; the comparison is an empirical model fit, not a mathematical consequence. The abstract's claim of 'improved model fit' is undermined by the paper's own Table 2, where WAIC favors the Euclidean R3 model (161.099) over the spherical S2 model (199.367), and Table 4's 'out-of-sample' predictive protocol is not described (no split, folds, or edge-sampling details). These are significant evidential and reproducibility concerns, but they do not constitute circularity: the predictive metrics are not shown to be identical to the hyperparameter selection criterion, and no equation or fitted parameter is renamed as a prediction. Self-citations (e.g., Sosa & Betancourt 2022 for centrality measures) are used for definitions and reviews, not as load-bearing justifications for the central claim. Therefore no circular step can be established from the manuscript text.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities; latent positions are inherited from Euclidean latent space models and the sphere is a modeling constraint rather than an invented object. The main ledger items are the fitted intercept, similarity scaling, latent positions, and data-dependent hyperparameters.

free parameters (5)

alpha (intercept) = estimated, not reported
Baseline log odds of a tie when latent vectors are orthogonal; estimated by MCMC and maximum likelihood in Section 4.
beta (similarity scaling) = estimated, not reported
Controls how strongly angular similarity influences tie probability; estimated jointly with alpha and Z.
Latent positions Z = n(d-1) values per model
The main fitted quantities: each actor receives a position on the sphere, estimated by MCMC.
Euclidean hyperparameters sigma_z and sigma_alpha = 5.0, 5.0
Fixed in Section 5 and selected via sensitivity analysis; they affect the Euclidean posterior and model comparison.
Spherical hyperparameters rho, mu_alpha, sigma_alpha, mu_beta, sigma_beta = -0.5, 0.0, 1.0, 10.0, 5.0
Set in Section 5 after sensitivity analysis to optimize information and predictive criteria; these are hand-chosen inputs to the comparison.

assumptions (5)

domain assumption Dyads are conditionally independent given latent positions and parameters.
The likelihood in Section 3 factorizes as a product over i<j; this is a standard conditional independence assumption in latent space models.
standard math Jointly exchangeable random effects can be represented as h(z_i, z_j) with i.i.d. latent variables z_i.
Invoked in Section 2.4 through Hoff (2008), Hoover (1982), and Aldous (1985) to justify the latent position formulation.
domain assumption The spherical geometry with a symmetric inner-product link is adequate for the networks under study.
Central modeling assumption of Section 3; if the true latent structure is not spherical, the model is misspecified and the comparison is unfair to either geometry.
domain assumption Excluding the isolated Florentine family does not change the model comparison.
Section 5 drops one of 16 families with no marriage ties to avoid infinite distances; this selection affects both Euclidean and spherical fits but changes the dataset.
ad hoc to paper Hyperparameters selected via sensitivity analysis are valid for model comparison.
Section 5 states that hyperparameters were chosen to optimize information and predictive criteria; this is a data-dependent prior choice, not a principled elicitation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spherical latent space models for social network analysis." pith.science (2026). https://pith.science/paper/4LVWFKVC

@misc{pith2026250816556,
  author       = {Pith},
  title        = {Pith review of: Spherical latent space models for social network analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LVWFKVC}},
  note         = {Machine review of arXiv:2508.16556}
}

read the original abstract

This article introduces a spherical latent space model for social network analysis, embedding actors on a hypersphere rather than in Euclidean space as in standard latent space models. The spherical geometry facilitates the representation of transitive relationships and community structure, naturally captures cyclical patterns, and ensures bounded distances, thereby mitigating degeneracy issues common in traditional approaches. Bayesian inference is performed via Markov chain Monte Carlo methods to estimate both latent positions and other model parameters. The approach is demonstrated using two benchmark social network datasets, yielding improved model fit and interpretability relative to conventional latent space models.

Figures

Figures reproduced from arXiv: 2508.16556 by the authors.

**Figure 2.** Comparison of sampling results for the Rosenbrock distribution across different spaces [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

**Figure 3.** Visualization of the marriage network among Florentine families. [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Log-likelihood trajectories for all models obtained during MCMC sampling. [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]

**Figure 5.** Figure 5: Latent space representations for R 1 and S 1 models, showing each node’s posterior distribution and mean position, with edges indicating ties from the original network. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]

**Figure 6.** Figure 6: Clustering across latent space representations, showing number of communities [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]

**Figure 7.** Figure 7: Posterior predictive checks for all models using network-level statistics in the Florentine [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]

**Figure 8.** Figure 8: Receiver operating characteristic (ROC) curves for each model. 6 Discussion This paper introduces a spherical latent space model for social network analysis, extending Hoff et al. (2002) by embedding nodes on a hypersphere instead of in Euclidean space. This geometry …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Latent space models for networks with nodal multiplicative effects
stat.ME 2026-08 conditional novelty 5.0 of 10

Nodal multiplicative distance scaling improves generative flexibility and structural fit of latent space network models.

Reference graph

Works this paper leans on

9 extracted references · 8 canonical work pages · cited by 1 Pith paper

[1]

Al-Taie, M. Z. and Kadry, S. (2017).Python for graph and network analysis. Springer. Albert, J. et al. (2009).Bayesian computation with R, volume

work page 2017
[2]

Aldous, D

Springer. Aldous, D. (1985). Exchangeability and related topics. École d’été de probabilités de Saint-Flour, XIII—1983, 1117:1–198. Aliverti, E. and Russo, M. (2022). Stratified stochastic variational inference for high- dimensional network factor model.Journal of Computational and Graphical Statistics, 31(2):502–511. Austin, A., Linkletter, C., and Wu, Z...

arXiv 1985
[3]

Raftery, A

Cambridge University Press Cambridge, UK:. Raftery, A. E., Niu, X., Hoff, P. D., and Yeung, K. Y. (2012). Fast inference for the latent space network model using a case-control approximate likelihood.Journal of computational and graphical statistics, 21(4):901–919. Rastelli, R. and Friel, N. (2018). Optimal bayesian estimators for latent variable cluster ...

work page 2012
[7]

Sethuraman, J

Courier Dover Publications. Sethuraman, J. (1994). A constructive definition of dirichlet priors. Statistica sinica, pages 639–650. Sewell, D. K. and Chen, Y. (2015). Latent space models for dynamic networks.Journal of the American Statistical Association, 110(512):1646–1657. Sewell, D. K. and Chen, Y. (2016). Latent space models for dynamic networks with...

work page 1994
[20]

Hoff, P. D. (2009). Multiplicative latent factor models for description and prediction of social networks. Computational and mathematical organization theory, 15(4):261–272. Hoff, P. D. (2011). Hierarchical multilinear models for multiway data.Computational Statistics & Data Analysis, 55(1):530–543. Hoff, P. D. (2015). Multilinear tensor regression for lo...

work page 2009
[233]

and Murphy, T

Gollini, I. and Murphy, T. B. (2016). Joint modeling of multiple network views.Journal of Computational and Graphical Statistics, 25(1):246–265. Golub, G. H. and Van Loan, C. F. (2013).Matrix computations. JHU press. Guhaniyogi, R. and Rodriguez, A. (2020). Joint modeling of longitudinal relational data and exogenous variables.Bayesian Analysis, 15(2):477...

work page 2016
[1087]

Padgett, J. F. and Ansell, C. K. (1993). Robust action and the rise of the medici, 1400-

work page 1993
[1434]

Pósfai, M

American Journal of Sociology, 98(6):1259–1319. Pósfai, M. and Barabási, A.-L. (2016).Network science, volume

work page 2016

Show all 9 references

[2025]

and Calvo, M.-P

Sanz-Serna, J.-M. and Calvo, M.-P. (2018).Numerical hamiltonian problems, volume

2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Al-Taie, M. Z. and Kadry, S. (2017).Python for graph and network analysis. Springer. Albert, J. et al. (2009).Bayesian computation with R, volume

work page 2017

[2] [2]

Aldous, D

Springer. Aldous, D. (1985). Exchangeability and related topics. École d’été de probabilités de Saint-Flour, XIII—1983, 1117:1–198. Aliverti, E. and Russo, M. (2022). Stratified stochastic variational inference for high- dimensional network factor model.Journal of Computational and Graphical Statistics, 31(2):502–511. Austin, A., Linkletter, C., and Wu, Z...

arXiv 1985

[3] [3]

Raftery, A

Cambridge University Press Cambridge, UK:. Raftery, A. E., Niu, X., Hoff, P. D., and Yeung, K. Y. (2012). Fast inference for the latent space network model using a case-control approximate likelihood.Journal of computational and graphical statistics, 21(4):901–919. Rastelli, R. and Friel, N. (2018). Optimal bayesian estimators for latent variable cluster ...

work page 2012

[4] [7]

Sethuraman, J

Courier Dover Publications. Sethuraman, J. (1994). A constructive definition of dirichlet priors. Statistica sinica, pages 639–650. Sewell, D. K. and Chen, Y. (2015). Latent space models for dynamic networks.Journal of the American Statistical Association, 110(512):1646–1657. Sewell, D. K. and Chen, Y. (2016). Latent space models for dynamic networks with...

work page 1994

[5] [20]

Hoff, P. D. (2009). Multiplicative latent factor models for description and prediction of social networks. Computational and mathematical organization theory, 15(4):261–272. Hoff, P. D. (2011). Hierarchical multilinear models for multiway data.Computational Statistics & Data Analysis, 55(1):530–543. Hoff, P. D. (2015). Multilinear tensor regression for lo...

work page 2009

[6] [233]

and Murphy, T

Gollini, I. and Murphy, T. B. (2016). Joint modeling of multiple network views.Journal of Computational and Graphical Statistics, 25(1):246–265. Golub, G. H. and Van Loan, C. F. (2013).Matrix computations. JHU press. Guhaniyogi, R. and Rodriguez, A. (2020). Joint modeling of longitudinal relational data and exogenous variables.Bayesian Analysis, 15(2):477...

work page 2016

[7] [1087]

Padgett, J. F. and Ansell, C. K. (1993). Robust action and the rise of the medici, 1400-

work page 1993

[8] [1434]

Pósfai, M

American Journal of Sociology, 98(6):1259–1319. Pósfai, M. and Barabási, A.-L. (2016).Network science, volume

work page 2016