acceptodds
Under review as a conference paper at ICLR 2027

Inferring Latent Geometric Embeddings from Grouped Relational Data

Abstract

Graph embedding methods place nodes so that distance matches an observed graph. In many domains no such graph exists, and relatedness between entities must be inferred from repeated, grouped counts collected across heterogeneous contexts. We introduce a generative model in which relatedness is latent rather than observed. Entity positions on a product manifold determine a pairwise propensity, and an observation layer links it to overdispersed counts. Flattening the manifold and taking one observation per pair recovers a Euclidean latent-space model. We derive the variance ratio between this geometric estimate and a fit using only a pair's own counts, and show it depends on how precisely the pair's endpoints are determined. A sparse pair between well-determined entities is therefore estimated nearly as well as a data-rich one. The ratio involves the geometry only through its dimension, so it applies to flat and curved embeddings alike. The same quantities yield a combined estimator with no fitted parameters. We sample the embedding's posterior with a Riemannian extension of stochastic gradient Langevin dynamics, and because the likelihood is generative, an entity absent from training can be placed from its own evidence alone. On cancer co-mutation data, 5% of that evidence suffices to outpredict full-evidence baselines on two thirds of genes.

Primary area:
generative models
Keywords:
graph embeddings, hyperbolic embeddings, latent space models, bayesian inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3JAoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies latent relational inference when no graph is observed directly. Instead of treating each entity pair as one binary edge, it models grouped counts collected across heterogeneous contexts. Each entity is embedded on a product manifold whose factors may be Euclidean, hyperbolic, or spherical. A differentiable distance-based link function converts the geometry into a pairwise relatedness propensity, and a Beta-Binomial observation model accounts for context-specific rates and overdispersion. The authors derive a variance ratio comparing the shared geometric estimate with an estimate based only on a pair's own counts, and construct a parameter-free combination of the two estimates. Posterior inference is performed with a Riemannian stochastic gradient Langevin sampler, producing uncertainty summaries and allowing a new entity to be embedded from its own evidence with shared parameters frozen. Synthetic experiments and evaluations on TCGA gene co-mutation and protein-protein interaction data show improved prediction in dropped-pair and held-out-entity settings.

Critical strengths

1.The problem raised in this paper is quite significant: Graph structures do not exist directly in the real world; rather, they are often inferred from observed statistical data.

2.Using Riemannian geometry to model the intrinsic geometric structure of data is both reasonable and well-motivated.

3.The experiments on biological data demonstrated the effectiveness of the proposed method.

Critical weaknesses

1.The variance-ratio result relies on an isotropic-neighborhood assumption that may be unrealistic for many relational datasets. In real data, information can be strongly directional: entities may be connected through several distinct relation types, have highly anisotropic local structure.

2.The experiments are limited to synthetic data and two biological datasets with about 1,000 entities. Broader evaluation on benchmarks such as RelBench, including heterogeneous-table, missing-relation, foreign-key, and inductive entity prediction tasks, would better test generalization beyond biological co-occurrence data.

3.The paper should cite and discuss more closely related work on Riemannian geometry and product-manifold representations.

[1]Latent Graph Inference using Product Manifolds. ICLR 2023

[2] Toward a Unified Geometry Understanding: Riemannian Diffusion Framework for Graph Generation and Prediction. NeurIPS 2025

[3] Neural Latent Geometry Search: Product Manifold Inference via Gromov-Hausdorff-Informed Bayesian Optimization. NeurIPS 2023.

Questions for authors

See Weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3JAo

We thank the reviewer for a detailed and positive assessment and for three specific suggestions, each of which we have incorporated in the revision.

W1: The isotropy assumption. We agree that real neighborhoods are directional, and the main result does not depend on isotropy:

  • Proposition 1 does not assume isotropy. It is a matrix expression, (∇zipij)⊤Ii−1(∇zipij)(\nabla_{z_i} p_{ij})^\top \mathcal I_i^{-1} (\nabla_{z_i} p_{ij}) per endpoint, where Ii\mathcal I_i sums one rank-one term per neighbor in that neighbor's direction. Strongly directional or anisotropic structure enters through Ii\mathcal I_i. Corollary 1 adds isotropy only to reduce this to a scalar in the information shares. The revised Sec. 6 states this explicitly.
  • The exact variance assumes neither isotropy nor the block-diagonal approximation. The revision adds the exact joint-information variance and proves that Proposition 1 is within a factor of 2 above and 1/λ21/\lambda_2 below it, where λ2\lambda_2 is the spectral gap of the pair design (App. A.5).
  • We measured the isotropic reduction on the fitted embeddings (App. A.4). On TCGA it is the less accurate step: it departs from the exact variance by a factor of 0.37–3.4 for 98% of pairs. The combined estimator is insensitive to this. Replacing the scalar variance with the exact one moves the mean weight from 0.954 to 0.952 and changes held-out MAE and log-likelihood by at most 4e-5. On PPI the mean weight moves by 0.002 and held-out MAE by at most 2e-5.
  • Several relation types. Multiple contexts enter through the per-pair information Iij=∑sI(pij;nijs)\mathcal I_{ij} = \sum_s \mathcal I(p_{ij}; n_{ijs}). Remark 1 states the condition that matters when structure is directional: positions are identified only if each entity's neighbors span enough directions, and degree alone is not sufficient.

W2: Breadth and scale of evaluation, and RelBench. We added two real domains (Sec. 7.2):

  • Mitochondrial variants from 1000 Genomes Phase 3 (2,503 individuals, 26 populations). This is still biology, but one where the generating structure is known: inheritance without recombination makes it a single tree. There, at f≥0.25f \ge 0.25, H5\mathbb{H}^5 lowers dropped-pair error on every seed relative to R5\mathbb{R}^5 (mean 8–16%), and H5\mathbb{H}^5 with per-entity intercepts lowers it relative to a low-rank model with the same intercepts (mean 8–9%) (Table 5; App. B.3, Fig. 7).
  • MovieLens-1M, outside biology: movies are entities, users the grouped observations, and occupation the context. At N=2,500N = 2{,}500 movies (~3.1M candidate pairs), at every retention level, shared positions lower dropped-pair MAE relative to an offset-only control (by 21–24% on average) and by half relative to population shrinkage, and beat NCN on MAE and rank correlation (Table 6; App. B.4).

On RelBench: its tasks are supervised predictions over relational databases with observed foreign-key links and entity features, which is the observed-graph setting. Our model applies where a database yields grouped counts instead. MovieLens-1M is built this way, from a user–movie rating table with users as the grouped observations and occupation as the context, and the same construction applies to transaction tables such as those in RelBench. RelBench's own supervised tasks require features and labels our model does not use, so we see them as complementary to our setting. On scale: TCGA, PPI and mtDNA have N=1,000N = 1{,}000 entities (about 500,000 candidate pairs); MovieLens has N=2,500N = 2{,}500 (about 3.1M, roughly 6×). The per-iteration cost is linear in the number of candidate pair–context cells, and one MAP fit on MovieLens takes about 30 minutes on a single T4 GPU (App. C).

W3: Related work. We added all three papers (Sec. 2 and App. E):

  • [1] Latent graph inference with product manifolds (Borde et al., ICLR 2023) drops the observed graph, as we do. It samples edges from distances in a learned product manifold to improve a downstream node-level prediction. It requires node features and labels and treats the graph as an intermediate representation. Our model takes neither, and relatedness itself is the estimand.
  • [2] The Riemannian diffusion framework (GeoMancer; NeurIPS 2025) trains on observed graphs to generate new graphs or predict their properties, so it sits in the observed-graph setting.
  • [3] Neural latent geometry search (Borde et al., NeurIPS 2023) selects the product-manifold signature by Bayesian optimization over a downstream score. It complements our work: it could choose the signature on top of our likelihood. On our data, a mixed H2×R3\mathbb{H}^2 \times \mathbb{R}^3 signature did not improve on a single factor on TCGA or PPI (Sec. 7.2).

We hope the new domains, the exact-variance results and the product-manifold discussion address W1–W3. If so, we would be grateful if the reviewer would consider raising their score. We would be glad to discuss any of these points further.

Thanks for authors' reply. My concern has already been solved. I would like to keep my positive score unchanged.

Reviewer_PwGFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how to estimate relationships from noisy, repeated observations. For each entity pair, the data record how often an event occurs out of a given number of trials. These counts are grouped by context, such as cancer cohorts or experimental screens. Direct frequencies can be unreliable when observations are scarce. Estimating each pair separately also ignores information from other relationships.

The proposed method gives each entity a position in a latent geometric space. Distances between positions determine pairwise relationship probabilities. A Beta-Binomial observation model allows event rates to vary across contexts. The authors fit the model using MAP estimation and approximate posterior sampling. They also analyze when shared embeddings reduce estimation variance and propose combining geometric predictions with population shrinkage.

Experiments use synthetic data, TCGA gene co-mutation data, and protein-interaction screening data. They evaluate estimation with limited observations, prediction for missing pairs, and adaptation to new entities. Baselines include independent pairwise estimators, population shrinkage, a Euclidean latent-space model, and graph-embedding methods. The reported results show benefits from sharing information across pairs, while the gains from curved geometry and the combined estimator depend on the dataset.

Critical strengths
  1. Practical problem. Estimating relationships from sparse observations across heterogeneous contexts is relevant to applications such as gene co-mutation and protein-interaction analysis. The formulation retains sample-size and context information that can be lost when constructing a graph upfront.

  2. The novelty is not sufficiently established. The method combines established components, including latent-distance modeling, curved geometries, and a Beta-Binomial observation model. The paper should more clearly identify the substantive advances provided by its information-sharing analysis and combined estimator. The external baselines also primarily draw on methods from 2019 or earlier, leaving it unclear whether the comparisons adequately represent current approaches to this problem.

Critical weaknesses
  1. Writing and presentation are difficult to follow. The problem setting and motivation are not clearly introduced, making it difficult to identify the paper’s central contribution. The distinction between “graph embedding with a fixed, observed graph” and “latent space models for network data” is also confusing, since graph embeddings themselves provide latent representations. The paper should explain the modeling assumptions that distinguish these overlapping categories.

  2. Limited baseline coverage. The comparisons mainly involve classical pairwise estimators and older graph-embedding methods. While these are useful references, the paper does not sufficiently justify whether they represent the strongest relevant alternatives for this setting, limiting the conclusions that can be drawn about its comparative performance.

Questions for authors

My major concerns are listed in the weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer PwGF

We thank the reviewer for a clear summary and for specific feedback on the presentation, which we took seriously. We revised the sections that set up the problem and the main result; all changes are in the revised PDF.

1. Revisions to the presentation.

  • Graph embedding vs. latent space models (Sec. 2). Sec. 2 now states the modeling difference directly (point 3 below).
  • Problem setting (Sec. 3). Background now opens with a roadmap and a setting paragraph with a concrete example: genes as entities, cancer types as contexts, and kijsk_{ijs} the number of tumors of type ss in which both genes are mutated. No graph is observed, and the goal is a calibrated estimate of every pair's relatedness, including pairs with little or no data of their own.
  • Why shared positions help (Sec. 3.1). Because distances obey the triangle inequality, positions make relations transitive: two entities close to a common third are close to each other. This is the mechanism behind the variance results.
  • The central result (Sec. 6). The section now opens with this argument worked through. If ii and jj are both strongly related to kk, the embedding places both near kk, and hence near each other, predicting a high pijp_{ij} even when the pair is rarely observed. It then explains how to read the variance ratio: a ratio of 1/41/4 means the pair's own counts would need four times their information to match the shared estimate.

2. Novelty relative to known components. Latent-distance modeling, curved geometry and the Beta-Binomial are established, and we cite each. The advances are what we derive from combining them:

  • Which pairs benefit (Sec. 6, App. A). Classical shrinkage says pooling helps on average but not which pairs benefit. We show that a pair's gain is governed by its share of its endpoints' information, not by its own sample size, so a sparse pair between two well-determined entities is estimated nearly as well as a data-rich one. The revision adds the exact joint-information variance and bounds the error of Proposition 1 in both directions (App. A.5). As far as we know, no result of this kind exists in the shrinkage or latent space literature (App. A, Remark 2).
  • A combined estimator with no additional fitted parameters (Eq. 13, Sec. 6; derivation in App. A.6). The same per-pair quantities set the weight between the geometric estimate and population shrinkage. On PPI it improves on both estimates it combines: AUPRC is .198, against .126 for the geometric estimate and .158 for population shrinkage, and precision at 100 is .66 against .44 and .58 (Table 8).

3. Graph embedding vs. latent space models. Both produce latent representations; they differ in what is observed and whether the data are modeled. Graph embedding (Nickel & Kiela; Gu et al.) takes an observed graph as input and fits positions to its edges with a ranking or distortion loss, with no probability model of how the data arose. Latent space models (Hoff et al.) are generative models of an observed network, typically one binary tie per pair. Our model is generative, but no graph is observed: the data are overdispersed counts per pair and context. Sec. 2 now states this directly. To run graph-embedding baselines at all, we must first threshold the counts into a graph, which discards the trial counts and context structure.

4. Baselines, including post-2019 methods. We added two baselines:

  • Neural Common Neighbor (NCN; Wang et al., ICLR 2024), a recent graph-neural link predictor, trained on the same induced graph. Dropped-pair MAE on TCGA is .041–.043 against .037–.039 for our geometry-only fit (Table 3). On PPI its AUROC is .798 against .891 for our combined estimator (Table 8).
  • A low-rank eigenmodel (Hoff, 2008) with per-entity intercepts under our Beta-Binomial likelihood, the strongest low-rank model fit to the same data. With matched intercepts, H5\mathbb{H}^5 ties it on TCGA through f=0.1f = 0.1 and is lower at f=0.05f = 0.05, and is lower on mtDNA on every seed (mean 8–13%).

More recent graph methods need an observed graph, node features, or both, and our data have neither. Latent graph inference on product manifolds (Borde et al., ICLR 2023), for example, learns a graph from node features to improve a downstream node-level prediction (Sec. 2, App. E).

We also added mitochondrial variants from 1000 Genomes, whose generating structure is a tree; at every retention level where the comparison is informative (f≥0.25f \ge 0.25), curvature lowers error against every matched flat model (Sec. 7.2, Table 5). We also added MovieLens-1M, outside biology and at 2.5× the scale (N=2,500N = 2{,}500 movies, ~3.1M candidate pairs), where shared positions halve dropped-pair MAE relative to population shrinkage and beat NCN on MAE and rank correlation (Sec. 7.2, Table 6). If the rewritten sections make the contribution clearer, we would be grateful if the reviewer would consider revising their score. We would be glad to clarify further.

Thanks for authors' reply. My concern has already been solved. Raised the score.

Reviewer_myXiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers the problem of estimating pairwise event probabilities from counts observed across different contexts. It proposes a generative model in which each entity has a geometric embedding, pairwise probabilities are determined by distances between embeddings, and observed counts follow a Beta-Binomial distribution. The authors derive an approximate expression for the variance reduction obtained by using shared embeddings rather than estimating each pair independently. They also use posterior sampling to estimate uncertainty in the embeddings and evaluate the method on synthetic data, cancer co-mutation data, and protein interaction data, reporting improvements over the baselines considered.

Critical strengths

S1 — Variance analysis

The paper attempts to quantify the reduction in estimation variance obtained by using shared embeddings rather than estimating each pair independently. The analysis relates this reduction to how well the embeddings of the two entities are determined by observations involving other pairs. This is a useful way to understand when sharing information through embeddings is beneficial.

S2 — Empirical results

The method shows improvements over the baselines considered on cancer co-mutation and protein interaction datasets. In particular, on protein interaction prediction, the combined estimator improves AUPRC from 0.158 to 0.198 and precision@100 from 0.580 to 0.660 compared to population shrinkage.

Critical weaknesses

W1 — Limitations of the variance analysis

The variance analysis assumes that the uncertainty in each embedding can be estimated while treating the other embeddings as fixed. However, the embeddings are estimated jointly, and the resulting dependencies are not accounted for. Although the authors acknowledge this approximation in Appendix A.2, they do not establish conditions under which it is accurate or quantify the approximation error. This limits the strength of the paper's main theoretical contribution.

W2 — Baselines and experimental validation

The paper compares against several established baselines, but it is unclear whether these represent the state of the art for this estimation problem. Further, the synthetic experiments generate data using the assumed geometric model, while the real-data experiments show limited benefits from non-Euclidean geometries. This leaves the practical importance of the geometric assumptions unclear.

Questions for authors

Q1: Can the authors provide conditions under which the variance approximation in Proposition 1 accurately captures the variance of the jointly estimated embeddings? An empirical comparison with the variance obtained from the full joint information matrix would help assess the accuracy of this approximation.

Q2: How were the baselines selected, and do they represent the strongest available methods for this problem? In particular, how does the proposed method compare with standard low-rank probabilistic models using the same Beta-Binomial likelihood?

2 comments
AuthorsOpenReview ↗

Response to Reviewer myXi (1 of 2)

We thank the reviewer for a close reading of the theory, including App. A.2, and for the suggestion to compare against the full joint information matrix. We have now done the comparison by computing the full joint information and proving two-sided bounds. Proposition 1 is within 10% of the exact variance for ≥99% of pairs on the transductive real-data fits. It understates the variance for dropped pairs as data thin out, and a closed-form second-order term recovers most of the gap (App. A.5, Tables 1–2).

W1 / Q1: Conditions for the block-diagonal approximation, and comparison with the full joint information.

App. A.5 computes the exact quantity that Proposition 1 approximates. The information over all NN positions is J=∑i<jIijGijGij⊤\mathcal{J} = \sum_{i<j} \mathcal I_{ij} G_{ij} G_{ij}^\top, where Gij=∇zpijG_{ij} = \nabla_z p_{ij}. The likelihood is invariant under the isometry group, so J\mathcal{J} has a kernel of dimension at least D(D+1)/2D(D+1)/2. Each pijp_{ij} is itself invariant, so the exact delta-method variance Vjoint=Gij⊤J+GijV^{\mathrm{joint}} = G_{ij}^\top \mathcal{J}^{+} G_{ij} is well defined, where J+\mathcal{J}^{+} is the Moore–Penrose pseudo-inverse (needed because J\mathcal{J} is singular along the isometry directions; since GijG_{ij} is orthogonal to that kernel, the value does not depend on the choice of generalized inverse). We prove two results:

  • Precision additivity (Prop. 2). 1/Vjoint=1/V−ij+Iij1/V^{\mathrm{joint}} = 1/V_{-ij} + \mathcal I_{ij}, where V−ijV_{-ij} is the variance of pijp_{ij} inferred from every pair except (i,j)(i,j). The exact variance ratio is therefore ρ/(1+ρ)≤1\rho/(1+\rho) \le 1, with ρ=IijV−ij\rho = \mathcal I_{ij} V_{-ij}. V−ijV_{-ij} is also the variance that governs the dropped-pair experiments.
  • Two-sided bound (Prop. 3). 12Vprop≤Vjoint≤Vprop/λ2\tfrac{1}{2} V^{\mathrm{prop}} \le V^{\mathrm{joint}} \le V^{\mathrm{prop}}/\lambda_2, where VpropV^{\mathrm{prop}} is Proposition 1 and λ2\lambda_2 is the smallest nonzero eigenvalue of D−1/2JD−1/2\mathcal{D}^{-1/2}\mathcal{J}\mathcal{D}^{-1/2}, with D\mathcal{D} the block-diagonal part of J\mathcal{J} (the per-node matrices Ii\mathcal I_i that Proposition 1 uses).

This gives the conditions the reviewer asked for. Proposition 1 is accurate when the observed-pair framework is rigid (the kernel contains only the isometries) and well connected (λ2\lambda_2 is not small). It can fail when removing a pair leaves the framework flexible in a direction that changes dijd_{ij}, even if both endpoints are individually well determined. App. A.2 now gives the first two correction terms in closed form. The first, −2Iijbibj-2\mathcal I_{ij} b_i b_j, is negative because the pair is counted at both endpoints. The second is non-negative, because the endpoints' neighbors are themselves estimated. For a dropped pair the first vanishes, so Proposition 1 understates at leading order. On the TCGA dropped-pair fits, adding the second term lowers the median error from up to 17% to at most 6% (Table 1); on PPI, adding both terms lowers it from 2.6% to 0.4%.

Empirical comparison. We assembled the full 5000×50005000 \times 5000 matrix J\mathcal{J} for the fitted H5\mathbb{H}^5 embeddings on TCGA and PPI.

  • Transductive fits (Table 2). The kernel has dimension exactly 15=dim⁡so(5,1)15 = \dim \mathfrak{so}(5,1) on every fit, so the framework is rigid. Vjoint/VpropV^{\mathrm{joint}}/V^{\mathrm{prop}} is within 10% of 1 for 100% of TCGA pairs and 99.3% of PPI pairs (λ2\lambda_2 = 0.84–0.85 on TCGA, 0.11–0.15 on PPI). The error tracks the largest information share relative to λ2\lambda_2. That quantity is 0.006 on TCGA, where 98% of pairs are within 0.7% of the exact variance, and 1.7–2.4 on PPI, where the same range is −2% to +9%.
  • TCGA dropped-pair fits (Table 1). Proposition 1 understates the leave-pair-out variance more as data thin out. The median Vprop/V−ijV^{\mathrm{prop}}/V_{-ij} is 0.98–0.99 at retention f=0.5f = 0.5 and 0.83–0.89 at f=0.05f = 0.05. Holding the other positions fixed omits the uncertainty in each endpoint's neighbors, which grows as their own pairs are removed.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer myXi (2/2)

Q2: A low-rank model with the same Beta-Binomial likelihood.

We added exactly this comparison: the eigenmodel of Hoff (2008), pij=σ((ai+aj+∑dλduidujd)/t)p_{ij} = \sigma\big((a_i + a_j + \sum_d \lambda_d u_{id} u_{jd})/t\big). It uses the same Beta-Binomial likelihood, priors, seeds, splits and retention levels as our model. Since −∥ui−uj∥2=2⟨ui,uj⟩−∥ui∥2−∥uj∥2-\lVert u_i - u_j \rVert^2 = 2\langle u_i, u_j \rangle - \lVert u_i \rVert^2 - \lVert u_j \rVert^2, an eigenmodel with free intercepts contains the Euclidean distance model. The matched comparison is therefore against our H5\mathbb{H}^5 model with the same per-entity intercepts. Dropped-pair MAE, mean over seeds or splits:

f=0.5f=0.5f=0.25f=0.25f=0.1f=0.1f=0.05f=0.05
TCGA, eigenmodel + intercepts.0360.0361.0371.0452
TCGA, H5\mathbb{H}^5 + intercepts.0360.0361.0369.0413
mtDNA, eigenmodel + intercepts.0110.0124——
mtDNA, H5\mathbb{H}^5 + intercepts.0101.0113——

On TCGA the two tie through f=0.1f = 0.1; at f=0.05f = 0.05 both intercept models degrade (each is worse than H5\mathbb{H}^5 without intercepts, .0385), the eigenmodel more. On mtDNA, H5\mathbb{H}^5 with intercepts is lower on every seed, by 13% on average on retained pairs (f=1f = 1) and 8–9% at f=0.5f = 0.5 and 0.250.25. Below f=0.25f = 0.25 no model beats predicting zero on mtDNA (89% of held-out pairs never co-occur), so we report those levels in Table 5 but do not use them to compare geometries. On PPI, every fitted λd\lambda_d is positive, so the eigenmodel is the R5\mathbb{R}^5 distance model with intercepts. It matches H5\mathbb{H}^5 with intercepts (AUROC .898 vs .899, AUPRC .186 vs .194, P@100 .663 vs .690).

How baselines were selected (Q2). Each baseline isolates one ingredient of the model:

  • per-pair counts with pooling toward a global rate (population shrinkage, the standard empirical-Bayes treatment of co-occurrence counts)
  • the domain's standard co-occurrence method (DISCOVER, on TCGA)
  • the flat single-observation latent space model our model reduces to (Hoff)
  • graph embeddings given an induced graph (N&K, Gu)
  • the same-likelihood low-rank model above

We also added a recent graph-neural link predictor, Neural Common Neighbor (NCN; Wang et al., ICLR 2024), trained on the same induced graph. Its dropped-pair MAE is .041–.043 on TCGA, against .037–.039 for our geometry-only fit. On mtDNA it is .023–.024, which is worse than predicting zero. On PPI its AUROC is .798 and P@100 is .13, against .891 and .66 for our combined estimator. These methods take a binary graph as input, and thresholding the counts into one discards the trial counts and context heterogeneity that our likelihood models.

W2: Synthetic data from the assumed model; value of curvature on real data. The synthetic experiments use the assumed model so that the true rates are known and Corollary 1 can be checked against them, which it matches to within 3% (Fig. 2C). On real data we agree that on TCGA and PPI curvature adds little. There, the gain from sharing positions holds in flat and curved geometry alike, as Proposition 1 predicts, since it contains no curvature term. To test curvature where the generating structure is known, we added a third real domain: mitochondrial variants from 1000 Genomes Phase 3 (2,503 individuals, 26 populations, the 1,000 most frequent alternate alleles). Mitochondrial DNA is inherited without recombination, so carrier patterns arise from a single tree. Under the same likelihood, H5\mathbb{H}^5 lowers MAE relative to R5\mathbb{R}^5 by 23% on retained pairs (f=1f = 1) and by 16% and 8% on dropped pairs at f=0.5f = 0.5 and 0.250.25, on every seed (Table 5; App. B.3, Fig. 7). The intercept comparison above shows that the gain is not degree heterogeneity alone.

If the joint-information analysis and the matched eigenmodel comparison address Q1 and Q2, we would be grateful if the reviewer would consider updating their assessment. We would be glad to clarify anything further.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and summary of revisions

We thank the reviewers for careful and constructive reviews. They found the problem addressed in this paper "quite significant" (3JAo), since graphs are usually inferred from statistical data rather than observed, and noted that our formulation "retains sample-size and context information that can be lost when constructing a graph upfront" (PwGF). They described the variance analysis as "a useful way to understand when sharing information through embeddings is beneficial" (myXi) and the use of Riemannian geometry as "both reasonable and well-motivated" (3JAo), and found clear improvements on the biological datasets (myXi, 3JAo). The theoretical questions concern how accurate the block-diagonal approximation is (myXi) and how restrictive the isotropy assumption is (3JAo), and we answer both directly.

The questions fall into four groups:

  1. Accuracy of the theory. How good is the block-diagonal approximation in Proposition 1 relative to the full joint information (myXi), and how restrictive is the isotropy assumption (3JAo)?
  2. Baselines. Are the baselines the strongest available, including a low-rank model under the same Beta-Binomial likelihood (myXi) and methods more recent than 2019 (PwGF)?
  3. Value of curvature and breadth of evaluation. Real-data gains from non-Euclidean geometry were limited (myXi), and the evaluation covered two biological domains of about 1,000 entities each (3JAo).
  4. Positioning. What is new relative to known components, and how do graph embedding and latent space models differ (PwGF)? Which recent product-manifold work is related (3JAo)?

Summary of revisions (all in the revised PDF):

  • Exact joint-information variance (App. A.5, Tables 1–2). We derive the exact delta-method variance from the Fisher information over all positions and prove two results: the exact variance ratio never exceeds 1 (precision additivity), and Proposition 1 is within a factor of 2 above and 1/λ21/\lambda_2 below the exact variance, where λ2\lambda_2 is the spectral gap of the pair design. We compute it on the fitted H5\mathbb{H}^5 embeddings for TCGA and PPI, including the TCGA dropped-pair fits. Sec. 6 now states these results.
  • Isotropy. Proposition 1 is a matrix expression and does not assume isotropy; only the scalar reduction in Corollary 1 does (Sec. 6 now says this). On the fitted embeddings, replacing the scalar form with the exact variance barely changes the combined estimator: on both TCGA and PPI the mean weight moves by at most 0.003 and held-out MAE by at most 4e-5 (App. A.4).
  • A same-likelihood low-rank baseline. We add a low-rank eigenmodel (Hoff, 2008) with per-entity intercepts, under our Beta-Binomial likelihood, priors, seeds and splits. With matched intercepts it ties H5\mathbb{H}^5 on PPI and MovieLens and on TCGA through f=0.1f = 0.1, as Proposition 1 predicts, since the gain from sharing positions has no curvature term; at the sparsest TCGA level (f=0.05f = 0.05) both intercept models degrade, the eigenmodel more (.0452 vs .0413). On mtDNA, where the generating structure is a tree, H5\mathbb{H}^5 with the same intercepts is lower on every seed (mean 8–9% on dropped pairs; Tables 3–6, 8).
  • A recent baseline. We add Neural Common Neighbor (NCN; ICLR 2024) on all four real domains. Our model has lower MAE on every domain: dropped-pair MAE on TCGA is .037–.039 against .041–.043 for NCN, and AUROC on PPI is .891 against .798.
  • Two additional real domains (Sec. 7.2).
    • Mitochondrial variants from 1000 Genomes Phase 3, where inheritance without recombination makes the generating structure a single tree. H5\mathbb{H}^5 lowers dropped-pair MAE relative to R5\mathbb{R}^5 on every seed at f≥0.25f \ge 0.25 (mean 8–16%; Table 5; App. B.3, Fig. 7).
    • MovieLens-1M, outside biology and at 2.5× the scale (N=2,500N = 2{,}500 movies, ~3.1M candidate pairs). At every retention level, shared positions lower dropped-pair MAE relative to an offset-only control (by 21–24% on average) and by half relative to population shrinkage, and beat NCN on MAE and rank correlation (Table 6; App. B.4).
  • Related work. We discuss the three papers 3JAo suggested and state where each sits relative to our setting.
  • Presentation. Sec. 2 now states how graph embedding and latent space models differ. Sec. 3 opens with the setting and the estimand (grouped counts across contexts, no observed graph, a calibrated propensity for every pair) and explains why shared positions help: distances obey the triangle inequality, so entities close to a common third are close to each other. Sec. 6 opens with a worked example of this intuition, and the comparison methods are introduced once, grouped, at the start of Sec. 7.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.