acceptodds
Under review as a conference paper at ICLR 2027

Bayesian Matrix-Valued Graphs for Context-Dependent Multivariate Relationships

Abstract

Many graph-based models involve nodes with multivariate feature vectors, so scalar edge weights are insufficient for direction-dependent interactions. We develop the Bayesian matrix-weighted graph (BMWG), which assigns each edge a symmetric positive-definite matrix \(W_e\) and infers its context-dependent posterior geometry. The geodesic distance induced by the affine-invariant Riemannian metric (AIRM) quantifies the magnitude of change, and generalized eigenvalues resolve its signed directions of strengthening and weakening. We show that BMWG achieves competitive recovery of the full graph precision matrix against fused graphical lasso, Bayesian multiple-GGM, and common principal components, while retaining identifiable matrix-weighted edges and more accurately recovering edge-level deformation directions. In Bay Area Meteostat weather data, 12-hour evolution reconfigures spatial coupling by an amount comparable to seasonal variation. In TCGA-BRCA, estrogen-receptor-associated reconfiguration concentrates on specific gene-module pairs and persists under scaffold sparsification and removal of subgroup mean differences. BMWG therefore provides a unified framework for quantifying and interpreting context-dependent multivariate reconfiguration.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian inference, matrix-valued graphical models, Riemannian geometry, structured precision matrices, multivariate representation learning, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_9QasOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents Bayesian matrix-weighted graph (BMWG), a framework that improve edge weight representation for multivariate graph nodes, by encoding it as an SPD rather than a scalar, which allows representation of conditional variable effects.

Critical strengths
  • An expressive framework that goes beyond scalar representation in weighted edge to represent entire classes of SPD matrices
  • A method of high relevance to multiple domains in graph theoretical applications Demonstrated insights in 2 use case applications
  • Method rests on validated body of work in GGMs and robust sampling of SPD posteriors
Critical weaknesses
  • Not clear if the method can scale up to many thousands of variables like in biology and that’s what determines real world usability
  • A few passages are really hard to read for a human like “We claim no new SPD metric or generic MCMC method. Our contribution is their graph-specific synthesis into a posterior over matrix-weighted edges”, not sure who “their” referes to. Significant rewriting is needed to make it appealing to readers
  • The cancer example is hardly relevant because it is only a 5-gene system, I would drop it for more expansive benchmarks against similar work such Multi-attribute GGMs, and Laplacian-constrained GGMs work
Questions for authors
  • It is not clear how multivariate data enter the likelihood estimate. Do sample estimate collapse to a single number per covariate or per data point?
  • How does runtime and how cost scales with nodes and d?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9Qas

We thank the reviewer for the positive evaluation.

  1. Q1 (how data enter the likelihood): Node features are never reduced to a single number before modeling. In the incidence model, each observation is the stacked vector Yr∈RmdY_r\in\mathbb R^{md} of all node features, the likelihood is Gaussian with precision Q(W)=R+∑eBeWeBe⊤Q(W)=R+\sum_eB_eW_eB_e^\top, and it depends on the data through the md×mdmd\times md sample second moment (Section 3, Appendix A.1). Each edge contributes the term (yi−yj)⊤We(yi−yj)=∑k,l(We)kl(yi−yj)k(yi−yj)l(y_i-y_j)^\top W_e(y_i-y_j)=\sum_{k,l}(W_e)_{kl}(y_i-y_j)_k(y_i-y_j)_l. The term is a scalar, as every log-likelihood term is, but it weights every pair of variables through the full matrix WeW_e, whereas a scalar edge weight wew_e would give only we∥yi−yj∥2w_e\|y_i-y_j\|^2, the same weight in every direction. In TCGA, the fixed edge-specific CCA maps yield scalar edge weights, but the Gaussian likelihood still acts on the 15-dimensional observation vector (Appendix A.6).

  2. W1/Q2 (runtime): Appendix C.1 gives the dense per-evaluation cost O(p3+p2qd+qd3)O(p^3+p^2qd+qd^3), where p=mdp=md and q=∣E∣q=|E|, plus the cost of pairwise geometric comparisons. In 25 end-to-end experiments on cycle graphs with m∈{5,10,20,50,100}m\in\{5,10,20,50,100\} nodes, mean workflow time rises from 2.61 to 80.82 minutes and ESS per second falls from 175.07 to 2.03 as mm grows from 5 to 100. All tested conditions pass the monitored convergence gates, but these experiments do not establish scalability to thousands of nodes or high-dimensional edge blocks. We have added hardware, software, and timing details. Since Q(W)Q(W) has the sparsity pattern of the graph, sparse Cholesky factorization is the natural route, which we note in the Discussion.

  3. W2 (readability): Thank you; "their" was ambiguous. We rewrote that passage and others in plain language and defined abbreviations at first use.

  4. W3 (cancer example; benchmarks): The TCGA analysis uses 15 genes in five prespecified three-gene modules (ten module-pair edges), not 5 genes. It is a heterogeneous setting: each module contains different genes, so variables cannot be matched across nodes. The fixed-map extension (Appendix A.6) handles this case, and TCGA is the only study that exercises it, which is why we keep the example. Laplacian-constrained GGMs are the d=1d=1 special case of BMWG (Proposition 1), and multi-attribute GGMs estimate a single context without a cross-context summary, so our empirical comparisons use methods designed to estimate and compare dependence structure across several contexts: fused graphical lasso, Bayesian multiple-GGM, and common principal components. Both families are now discussed in Section 2 and Table 3.

We hope these revisions address your concerns, and we would be glad to clarify anything further during the discussion period.

Thanks for you work on this.

  • (runtime) I think a direct consequence of the runtime analysis would be to specify were the limit of the models are in terms of variable size which would help modelers for an application domain. Possibly including the complexity analysis in the main text.

  • (use case) Which follows directly from the previous comment is that a realistic cancer analysis with thousands of genes is then not feasible. I'd suggest for the author to consider removing the cancer use case application, because with 15 gene no meaningful insights can be gained.

Thank you for the follow-up.

  1. Runtime limits: We agree and have stated the tested range explicitly. The complexity analysis is in the main text (Section 3, Computational cost), which now also gives the measured limits. The cost is driven by p=mdp=md, the total number of node variables. On cycle graphs (two neighbors per node), our dense implementation on an Apple M3 MacBook Pro takes 7.5 ms per sampler step at p=150p=150 and 20 ms at p=300p=300, and the full three-context workflow takes 29 and 81 minutes (Table 4, Appendix C.1). The largest size we tested is p=300p=300; these results do not establish scalability beyond it. Because Q(W)Q(W) has the sparsity pattern of the graph, sparse Cholesky factorization is the natural route to larger problems.

  2. Cancer use case: We agree that an analysis of 15 genes cannot establish meaningful biological findings or demonstrate scalability to genome-wide data. We therefore propose to move the TCGA analysis out of the main text and remove it from the abstract and contributions. A shortened version would remain in the appendix strictly as a methodological illustration, without biological claims: it is the only test of the fixed-map extension (Appendix A.6), where different nodes contain different variables, a case that Meteostat, whose nodes share aligned variables, does not cover. Would this address your concern? We would appreciate your perspective.

We will upload the revised PDF, including the runtime changes above, once we hear your view on the use case.

Reviewer_7rEEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a Bayesian matrix-weighted graph model for comparing multivariate relationships across contexts. It uses geometric summaries to characterize changes in edge strength and direction, and evaluates the framework on synthetic data and two real-world applications.

Critical strengths
  • The model provides an interpretable way to capture directional changes in relationships that scalar edge weights may miss.
  • The synthetic experiments cover several types and magnitudes of changes. The TCGA analysis also includes useful checks on graph structure and subgroup mean effects.
Critical weaknesses
  • My main concern is the novelty of the model. Since the inference and geometric analysis mainly use established techniques, the model formulation appears to be the central contribution. However, its relationship to existing multivariate graphical models and matrix-weighted graph models is not sufficiently clear. The related work should explain more precisely what is new in this formulation.
  • The real-data evaluation is somewhat limited. The weather network contains only eight stations, while the TCGA analysis uses one-dimensional CCA projections and scalar edge weights. Thus, the evidence for the full matrix-valued model mainly comes from one small real network. Results on another one or two networks with multidimensional edge weights would strengthen the paper.
  • The comparisons do not clearly isolate the benefit of Bayesian inference. A MAP baseline using the same matrix-weighted model would help distinguish the value of posterior inference from that of the representation itself.
Questions for authors
  • What is the main modeling difference from the closest prior formulations, and what additional capability does it provide?
  • How much of the observed improvement comes from the matrix-valued representation versus Bayesian inference?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 7rEE

We thank the reviewer for stating the main concern clearly.

  1. W1/Q1 (novelty): The closest formulations are multi-attribute GGMs (vector nodes, unrestricted blocks, one context), Laplacian-constrained GGMs (scalar attractive edges, point estimates), and matrix-weighted Laplacians (weights given, not inferred). Proposition 1 provides an exact necessary and sufficient characterization of the BMWG precision family for every node dimension dd: for fixed RR, the matrix Q−RQ-R has negative-definite edge blocks, zero nonedge blocks, and zero block-row sums. When RR is block diagonal, the resulting precision matrix has block-attractive conditional couplings; for d=1d=1 and diagonal RR, the model reduces to a Laplacian-constrained GGM on the fixed graph. Because each edge is an SPD matrix, two contexts can be compared geometrically: how much an edge changes, and which variable combinations strengthen or weaken, with posterior uncertainty and supporting theory (Theorems 2–3, Proposition 3). Section 2 and Table 3 state these differences explicitly.

  2. W3/Q2 (representation versus inference): We added the requested MAP baseline. Restricting edges to scalars (We=weIdW_e=w_eI_d, same model and sampler) increases candidate-edge error in 120/120 matched conditions (Appendix D.4), so the accuracy gains come from the matrix-valued representation. MAP and posterior means give similar point accuracy (Table 12); what Bayesian inference adds is uncertainty quantification: intervals for change magnitudes and signed directions, and the same-truth rule for deciding whether an edge changed. Theorem 2 guarantees that this MAP is the global optimum.

  3. W2 (real data): Meteostat is the real network with full matrix-valued edges; the TCGA study demonstrates the fixed-map extension (Appendix A.6) for heterogeneous data, comparing modules with different genes, and is robust to scaffold and mean-removal checks. We added a scalability study on graphs with up to 100 nodes and d=3d=3 (Table 4). We agree that further real networks with multidimensional edges would strengthen the evidence, and we state this in the Discussion.

We hope these revisions address your concerns, and we would be glad to clarify anything further during the discussion period.

Thank you to the authors for the additional clarifications and analyses addressing my concerns.

  1. Regarding my first concern, the authors have clarified the method’s position relative to existing work. One remaining question concerns its relationship to Dey (2026), particularly the model formulation and posterior inference discussed in Sections 2 and 5. It would be helpful to clarify the differences between the two studies and whether they should be considered concurrent work.
  2. My second concern is partially addressed. The additional synthetic scalability study is useful, but evidence for the full matrix-valued model on larger real-world networks remains limited.
  3. Regarding my third concern, the authors have provided the requested MAP comparison. The similar point-estimation accuracy suggests that the gains largely reflect the underlying matrix-valued model, with limited additional improvement from full posterior inference. I consider this concern addressed.

Overall, I acknowledge the substantive improvements and appreciate the authors’ efforts to address my concerns. Therefore, I decide to increase the rating.

Thank you for the follow-up and for raising your score.

  1. Relationship to Dey (2026): Dey (2026) is a non-peer-reviewed arXiv preprint that develops a sampler, ConeMALA (a Metropolis-adjusted Langevin method on the SPD cone), with a matrix-weighted Gaussian graph posterior as its target. We use ConeMALA only as the inference engine and do not claim it as a contribution. Our contributions are statistical: the exact model class and identifiability (Proposition 1, Theorem 1); the cross-context geometry of edge change and the same-truth reference (Section 4); posterior theory, including strict log-concavity with a unique MAP, a Bernstein-von Mises result, and misspecification (Theorems 2-3, Proposition 3); and the empirical comparisons and ablations (Section 5). We therefore view Dey (2026) as complementary work on computation; Section 2 now states this division.

The ablations confirm that the gains come from the model, not the sampler: Isotropic BMWG mixes well (split-R^≤1.0012\hat R\le1.0012, ESS ≥4439\ge4439) yet has higher candidate-edge error in all 120 matched conditions (Section 5.2), and the MAP, which uses no sampler, matches the accuracy of posterior means (Table 12).

  1. Real-world evidence: We agree; the Limitations (Section 6) now state that larger real networks remain to be studied.

  2. MAP comparison: We agree. As Theorem 3(a) predicts, MAP estimates and posterior means share the same limit; the posterior's role is uncertainty quantification. Both the same-truth detection rule and the intervals for signed directions rely on posterior draws.

We hope these revisions address your concerns, and we would be happy to incorporate any further suggestions within the discussion period.

Reviewer_Eij8OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Bayesian Matrix-Weighted Graphs, which is a probabilistic framework that represents graph edges using symmetric positive-definite matrices rather than the commonly used framework of scalar weights, thus enabling the characterization of direction-dependent multivariate relationships. To quantify changes in edge relationships across different contexts, it leverages ideas from Bayesian statistics and Riemannian geometry. The proposed framework is evaluated through experiments on synthetic data and two real world applications, one on meteorological data and one on breast cancer gene expression data.

Critical strengths
  • The use of symmetric positive-definite matrices to represent multivariate edge relationships is a good extension of the traditional scale edge weight framework. The formulation is clear and the paper provides results on identifiability. An additional advantage of the framework, is that it allows characterization of both changes in magnitude of edge relationships, as well as directions along which these changes occur, unlike the traditional framework.

  • The experimental results based on synthetic data support the proposed framework effectiveness. Further, the real data applications illustrate the usefulness of the framework in real domains.

Critical weaknesses
  • All the technical "ingredients" are already available in the literature and the key contribution is their integration in the proposed framework. While tis integration can informative in selected settings, it is unclear what key advancements it contributes to the large body of existing literature on multivariate graphical models, covariance estimation methods and Bayesian network analysis methods. In particular, the identification results are fairly standard, given the selected parametrization.

  • The experiments based on synthetic data generate data directly from the proposed underlying model and hence it has an inherent advantage over competing methods at evaluation time. It would be particularly useful, if the authors conduct experiment under mild and severe model misspecification.

  • A discussion on computational cost and scalability is largely missing. This is a concern, given that the experiments are on very small graphs, which is also true for the real world data sets.

  • The literature review is somewhat limited. There is broader body of literature and it would be helpful if a broader perspective was used to place the proposed methodology in context.

Questions for authors
  • Address computational and scalability issues.

  • Address framework performance under model misspecification. Incorporate more comparisons with competing methods.

  • How sensitive is the proposed framework to fixed graph scaffold, prior distributions and regularization parameters?

  • Given the somewhat limited scope of the literature review, expand and put the proposed methodology in context.

  • What do we learn from the real TCGA data, compared to competing approaches. Is there any biological significance?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Eij8

We thank the reviewer for the positive assessment and concrete suggestions.

  1. W1 (novelty; standard identifiability). We agree the ingredients are established. What is new (Section 2, Table 3): (i) the exact model class, block-attractive GMRFs with Qij=−We≺0Q_{ij}=-W_e\prec0, the vector-node analogue of Laplacian-constrained GGMs (Proposition 1); (ii) a signed, multivariate decomposition of edge change between contexts (AIRM magnitude plus generalized eigen-directions), with posterior uncertainty, which none of the related families provides; (iii) posterior theory (Theorems 2–3, Proposition 3). Identifiability is elementary once the parameterization is fixed; its value is that it holds for every simple graph and every dd without a rank condition, and Corollary 3 extends it to a learned block-diagonal stabilizer RR.

  2. W2/Q2 (misspecification). We added theory and simulations. Proposition 3: when the data are non-Gaussian or not block-attractive, BMWG estimates the closest model W†W^\dagger (KL projection onto the closed cone); the MAP converges to it with sandwich covariance, and the posterior concentrates at it with covariance I(W†)−1/n\mathcal I(W^\dagger)^{-1}/n. Appendix A.11 (Table 2, Figure 3) tests this:

  • Heavy tails (t5t_5): estimates stay correct, but credible intervals cover 68% instead of 90%; the sandwich correction restores 89%.
  • Missing edge: BMWG estimates the closest model on the given scaffold, with calibrated intervals (90–91%).
  • Repulsive coupling: the edge's smallest eigenvalue goes to 0 at rate 1/n1/n, so repulsion is flagged rather than turned into a misleading attractive edge.
  1. W3/Q1 (scalability). Each posterior evaluation costs O(p3+p2qd+qd3)O(p^3+p^2qd+qd^3) with p=mdp=md, q=∣E∣q=|E| (Section 3). On cycle graphs with up to m=100m=100 nodes (d=3d=3), BMWG ranks and detects the planted edge in all 25 experiments and flags none of the 900 unchanged edges; workflow time grows from 2.6 to 80.8 minutes as mm grows from 55 to 100100 (Section 5.1, Table 4).

  2. W4/Q4 (literature). See Section 2 and Table 3.

  3. Q3 (sensitivity). Scaffold: refitting TCGA on all 12 five-cycle sub-scaffolds gives edge distances with r=0.990r=0.990 to the full scaffold. Prior: signs of changing modes are stable across Wishart priors (Appendix D.6). Stabilizer: RR can be learned (Corollary 3).

  4. Q5 (TCGA). ER-associated reconfiguration is distributed but uneven: for six of ten module pairs, the 90% posterior interval for the log-ratio of edge weights excludes zero, so the direction of change (strengthening or weakening) is clear, and the largest change, on Structural/EMT–Signalling, is a roughly fivefold strengthening in ER-positive tumors (ηe∈[1.17,2.31]\eta_e\in[1.17,2.31]). This pattern is robust to scaffold sparsification and mean removal. We present the TCGA analysis as a demonstration on a prespecified 15-gene system, not as a biological discovery. We hope these revisions address your concerns, and we would be glad to clarify anything further during the discussion period.

Thank you again for your review. We would like to briefly note that the second revision addresses your points directly: the misspecification analysis (Proposition 3) with simulations (Appendix A.11), measured scalability up to p=300p=300 (Section 3, Table 4), and the positioning relative to multi-attribute and Laplacian-constrained GGMs (Section 2, Table 3). If any concern remains, we would be glad to clarify it before the discussion closes. If the revisions have addressed your concerns, we would be grateful if you would consider updating your assessment.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful reading and constructive suggestions. We uploaded a revised PDF in which substantive changes are shown in blue; we also revised the text throughout for clarity. The main changes are:

  1. Positioning (Section 2, Appendix B, Table 3): Related work is now organized around the closest families: multi-attribute GGMs (Kolar et al., 2014) and Laplacian-constrained GGMs (Egilmez et al., 2017), plus matrix-weighted, sheaf, and connection Laplacians.
  2. Exact model class (Proposition 1): For fixed R≻0R\succ0, a matrix QQ is a BMWG precision matrix if and only if Q−RQ-R satisfies the block-Laplacian conditions. When RR is block diagonal, the resulting Gaussian Markov random field has block-attractive conditional couplings. For d=1d=1 and diagonal RR, BMWG reduces to a Laplacian-constrained GGM on the fixed graph.
  3. Theory (Section 4.1, Appendices A.9–A.10): The posterior is strictly log-concave with a unique MAP (Theorem 2); a Bernstein–von Mises result explains why same-truth separation shrinks at rate n−1/2n^{-1/2} (Theorem 3); under misspecification, the posterior targets the closest model in KL divergence, and intervals need a sandwich correction (Proposition 3).
  4. Misspecification simulations (Appendix A.11). Heavy-tailed noise, a repulsive coupling, and a missing edge: all predictions of Proposition 3 hold.
  5. Cost and scalability (Sections 3 and 5.1, Appendix C.1). Graphs with up to 100 nodes.
  6. Representation versus inference (Section 5.2, Appendices D.4–D.6). Isotropic-BMWG and same-model MAP ablations, and a prior-sensitivity study.
AuthorsOpenReview ↗

General response: second revision

We have uploaded a second revision. Substantive changes since the previous revision are highlighted in blue; the previous revision, with the first-round changes in blue, remains available under Revisions. The main changes are:

  1. Following Reviewer 9Qas, the TCGA analysis has moved from the main text to Appendix F as a methodological case study of the fixed-map extension, and it has been removed from the abstract.
  2. Section 3 now reports measured run times and the largest tested size alongside the complexity analysis.
  3. Section 2 clarifies the relationship to Dey (2026), and Section 5.2 explains why the isotropic ablation separates the representation from sampling.
  4. The proofs in Appendices A.9-A.10 now give explicit references for each step.
  5. The abstract, Introduction, and Sections 3-6 were revised for readability.
AuthorsOpenReview ↗

Final revision uploaded

We sincerely thank all three reviewers and the Area Chair for their careful evaluation and constructive feedback. The discussion has helped us substantially strengthen the manuscript. We have clarified the distinct contributions of BMWG, expanded the theoretical results on identifiability, posterior inference, and model misspecification, and added experiments addressing comparisons, robustness, and computational scalability. We have also clarified the scope and limitations of the real-data applications. We have uploaded the final revision with all text in black. Compared with the previous (highlighted) revision, the only changes are minor ones from proofreading: consistent terminology and clarified definitions of evaluation metrics (e.g., “Axis error” in Section 5.2 and direction alignment in Table 8). No results have changed.

We believe these revisions have made the paper more rigorous, transparent, and complete. We appreciate the reviewers' time and consideration and hope the revised manuscript addresses their concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.