acceptodds
Under review as a conference paper at ICLR 2027

Nash-Guided Compatible Optimal Transport for Multimodal Knowledge Graph Fusion

Abstract

Multimodal Knowledge Graphs (MMKGs) integrate heterogeneous modalities to enhance knowledge representation. Yet, a fundamental challenge lies in multimodal representation fusion, where modality inputs often contain substantial target-irrelevant semantics (i.e., low-target relevance), and different modalities contribute unevenly during fusion (i.e., modality imbalance). Existing approaches typically align all modality representations indiscriminately and rely on pre-defined fusion strategies, making them sensitive to noisy or dominant modalities. To address these limitations, this paper proposes a Nash-guided Compatible Optimal Transport algorithm (NaCOT), which formulates multimodal fusion as an Optimal Transport (OT) problem. Specifically, NaCOT first constructs cascaded representation candidates and then employs sliced Wasserstein distance (SWD)-based selection to identify compatible source-target scales for OT alignment. Importantly, NaCOT introduces a Nash bargaining-based fusion mechanism that refines transport plans under fusion and OT constraints, jointly mitigating low-target relevance and modality imbalance. Experiments on four MMKG benchmarks show that NaCOT achieves state-of-the-art performance, with average relative improvements of 24.68% in MRR and 23.51% in Hit@1 over the strongest baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal knowledge graph, Optimal transport, Nash, Fusion

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_K68oOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an Optimal Transport based multimodal knowledge graph fusion framework, aiming at deal with the two critical problems about modality sematic noise ( low-target relevance ) and modality imbalance. The details of the solution are presented and the experiments on 4 popular benchmarks for performance evaluation and ablation are conducted to verify the effectiveness of the proposed method. The work is solid and significant for multimodal knowledge graph techniques.

Critical strengths

The two critical factors that individually and jointly affect the multimodality fusion result are analyzed carefully and the corresponding solutions are proposed. The cascaded compact multi-scale representation construction mechanism is designed to solve low-target relevance problem, the sliced Wasserstein distance based selection mechanism is designed to solve modality imbalance problem, and the Nash bargaining-based adaptive mechanism is designed to solve the joint problem. Finally the theoretical analysis and experimental results are given. The work is done completely and solidly.

However, the CRC and SWD mechanisms are designed by using the existing techniques such as cascaded factorization and sliced Wasserstein distance. Whether they are the suitable ones to solve the problems are to be explored further.

Critical weaknesses

None

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer K68o

We sincerely thank the reviewer for carefully reading our paper and providing such a positive assessment. We also appreciate the thoughtful question about whether cascaded factorization and sliced Wasserstein distance are suitable for addressing the identified problems. Our contribution lies in jointly adapting multi-scale candidate construction and compatibility-based selection to determine the representations used for OT-based MMKG fusion.

CRC compresses sample representations through learnable linear combinations. This reduces the OT problem size and computational cost (Appendix F), while aiming to mitigate the influence of weakly relevant information on OT alignment. Fig. 4 supports the benefit of moderate compression.

SCS selects the candidate pair with the smallest SWD. This favors closer distributional agreement and saves computation by avoiding exhaustive OT solves over all candidate pairs. Fig. 5 shows performance comparable to exhaustive Greedy selection.

The representation-level design retains information from partially relevant samples rather than explicitly filtering them out. A sample with substantial irrelevant content may still provide useful entity-specific evidence. CRC integrates such information through learnable linear combinations, while SCS selects a distributionally compatible pair for OT alignment. Appendix G.2 further shows that NaCOT maintains the highest MRR among the compared methods across all evaluated noise ratios on DB15K, supporting the robustness of the complete framework.

The existing techniques therefore serve specific, complementary roles in our design: CRC supplies compact alternatives to the original representations, and SCS identifies compatible pairs before OT alignment. The controlled comparisons in Figs. 4–5 support these choices. We have added this discussion and the supporting references to the revised manuscript.

Dear Reviewer K68o,

Thank you again for your encouraging assessment and thoughtful question about the suitability of CRC and SWD. We particularly appreciated that your question focused on why these techniques fit the problems we aim to address. It prompted us to explain their complementary roles more clearly and connect our design choices to the controlled comparisons and computational analysis. We have incorporated this discussion into the revised manuscript, making the motivation behind NaCOT more explicit.

We would welcome your thoughts on whether this explanation addresses your question, and would be happy to clarify any remaining points. Thank you again for your careful consideration and encouragement.

Best regards,
The Authors

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

This paper proposes NaCOT for representation fusion in multimodal knowledge graphs. It focuses on two main issues: text and images may contain information irrelevant to the target entity, and different modalities may contribute unevenly to the fusion process.

Critical strengths
  • Low-quality modality inputs and imbalanced modality fusion are both worthwhile problems to investigate.
  • Finding a suitable representation scale for alignment before adjusting modality fusion is a reasonable research direction.
  • The experimental evaluation is fairly extensive. It covers multiple knowledge graph completion datasets, as well as experiments involving noise, missing modalities, different encoders, and entity linking.
Critical weaknesses
  • The pseudocode appears to omit weight normalization, which leads to a different set of constraints. Specifically, Eq. (15) in the main text and line 15 of Algorithm 1 on page 33 do not seem consistent. The authors should clarify this discrepancy.

  • The model visualization is rather rough and could be improved to make the proposed method easier to understand.

  • On page 5, C_m is constructed from the "compressed candidate pairs," so its dimensions are the compressed source size multiplied by the compressed target size. However, the recovered P_m has dimensions n_m × N. Equation (12) still uses R(P_m) = <P_m, C_m> + τΩ(P_m), without redefining C_m at the full scale. Under the current definitions, this inner product is generally not well-defined.

  • The related work section does not adequately cover recent studies on multimodal knowledge graphs. The literature review should be updated to better reflect the latest developments in this area.

  • Some of the ablation experiments do not sufficiently isolate the contributions of individual modules, making it difficult to determine how much each component actually contributes to the final performance.

Questions for authors

see weakness

3 comments
AuthorsOpenReview ↗

[Part 1] Response to Reviewer ddyL

W1 The pseudocode appears to omit weight normalization, which leads to a different set of constraints. Specifically, Eq. (15) in the main text and line 15 of Algorithm 1 on page 33 do not seem consistent. The authors should clarify this discrepancy.

We thank the reviewer for pointing out this inconsistency. Eq. (15) in the main text is correct. The normalization factor ∑m∈Mωm(t−1)\sum_{m\in\mathcal{M}}\omega_m^{(t-1)} was inadvertently omitted from the consensus constraint in Line 15 of Algorithm 1. We have corrected the pseudocode to match Eq. (15).

W2 The model visualization is rather rough and could be improved to make the proposed method easier to understand.

We thank the reviewer for this valuable suggestion. We have redrawn Fig. 2 to provide a clearer and more detailed illustration of the NaCOT workflow. The revised figure explicitly presents the five main stages: (1) modality projection and cascaded representation construction (CRC), (2) SWD-based candidate pair selection (SCS), (3) initial optimal transport and transport plan recovery, (4) Nash-guided modality fusion (NMF), and (5) downstream KGC prediction.

In particular, we have added multi-scale representation diagrams to illustrate the CRC process, clarified how SCS selects compatible source–target representation pairs, and visualized the transition from compressed OT plans to full-size transport mappings. We have also expanded the NMF component to illustrate the iterative updates of transport plans, modality weights, and fused representations, including the Nash bargaining objective.

We have included the updated figure in the revised manuscript.

W3 On page 5, C_m is constructed from the "compressed candidate pairs," so its dimensions are the compressed source size multiplied by the compressed target size. However, the recovered P_m has dimensions n_m × N. Equation (12) still uses R(P_m) = <P_m, C_m> + τΩ(P_m), without redefining C_m at the full scale. Under the current definitions, this inner product is generally not well-defined.

We thank the reviewer for the careful observation. We inadvertently used CmC_m for both the compressed-scale and full-scale cost matrices. In the revised manuscript, we distinguish the full-scale matrix as C‾m∈Rnm×N\overline C_m\in\mathbb R^{n_m\times N}, defined by C‾m[i,j]=cm(Xm[i],Y[j])\overline C_m[i,j]=c_m(X_m[i],Y[j]). The regularization term in Eq. (12) is therefore written as R(Pm)=⟨Pm,C‾m⟩+τΩ(Pm).R(P_m)=\langle P_m,\overline C_m\rangle+\tau\Omega(P_m). We have corrected the notation and added this definition in the revised manuscript.

AuthorsOpenReview ↗

[Part 2] Response to Reviewer ddyL

W4 The related work section does not adequately cover recent studies on multimodal knowledge graphs. The literature review should be updated to better reflect the latest developments in this area.

We thank the reviewer for pointing out the need to incorporate more recent literature. We have revised the Related Work section by incorporating 11 papers published in 2026. Specifically, nine papers have been added to the MMKG discussion, covering recent advances in cross-modal interaction, structure-guided learning, adaptive multimodal fusion, and LLM-based KGC. We have also included one paper on OT-based cross-modal fusion and one on Nash bargaining to broaden the methodological context of NaCOT. The research focus of these additional papers is summarized in Table R1.

Table R1. The newly incorporated papers

CategoryPaper / MethodVenueMain Research Focus
MMKGM-HyperACL 2026Hypercomplex-driven multimodal collaboration and independence
MMKGCISEDEIEEE TNNLS 2026Cross-modal interaction through similarity enhancement and difference preservation
MMKGSECLIEEE TKDE 2026Structure-guided elastic consensus learning
MMKGTSAMIEEE TMM 2026Structure-aware multimodal representation learning
MMKGREFDIEEE TMM 2026Relation-aware entity feature decoupling
MMKGMDBGFFindings of ACL 2026Dynamic modality dropout and bidirectional gated fusion
MMKGMoCMEPattern Recognition 2026Complementarity-driven multimodal representation learning
MMKGHFR-MKGCAAAI 2026Hierarchical fusion and multimodal large language model reasoning
MMKGIT4MKGCInformation Fusion 2026Modality-aware instruction tuning for multimodal KGC
Optimal TransportCO²IFAAAI 2026Optimal transport-based cross-modal information fusion
Nash BargainingNBCGICML 2026Nash bargaining-based optimization for long-tailed multi-label learning

We have updated the Related Work section accordingly and thank the reviewer for this valuable suggestion.

[1] Liu et al. Collaboration of Fusion and Independence: Hypercomplex-driven Robust Multi-Modal Knowledge Graph Completion. ACL, 2026.

[2] Hao et al. Multimodal Knowledge Graph Completion by Cross-Modal Interaction With Similarity Enhancing and Difference Embracing. IEEE TNNLS, 2026.

[3] Liu et al. Evolution Rather Than Degradation: Structure-Guided Elastic Consensus Learning for Multimodal Knowledge Graph Completion. IEEE TKDE, 2026.

[4] Li et al. Towards Structure-Aware Model for Multi-Modal Knowledge Graph Completion. IEEE TMM, 2026.

[5] Hao et al. REFD: Relation-aware Entity Feature Decoupling for Multimodal Knowledge Graph Completion. IEEE TMM, 2026.

[6] Peng et al. Not All Modalities at Once: Dynamic Dropout and Bidirectional Fusion for Robust Multi-modal Knowledge Graph Completion. Findings of ACL, 2026.

[7] Li et al. Complementarity-driven Representation Learning for Multi-modal Knowledge Graph Completion. Pattern Recognition, 2026.

[8] Wang et al. HFR-MKGC: Hierarchical Fusion Reasoning with MLLMs for Multi-modal Knowledge Graph Completion. AAAI, 2026.

[9] Cui et al. Modality-aware Instruction Tuning of Large Language Models for Multimodal Knowledge Graph Completion. Information Fusion, 2026.

[10] Zhang et al. CO²IF: Language-Bridging Hyperspectral-Multispectral Image Fusion with Coordinated and Cross-modal Optimal Transport. AAAI, 2026.

[11] Yang et al. NBCG: Nash-Bargained Causal Game for Long-Tailed Multi-Label NLP. ICML, 2026.

W5 Some of the ablation experiments do not sufficiently isolate the contributions of individual modules, making it difficult to determine how much each component actually contributes to the final performance.

We thank the reviewer for raising this point. The Model Breakdown study (Lines 367–373, Fig. 3) compares three module-specific variants against the same standard OT baseline. Specifically, +CRC uses one cascade step (L=K=1L=K=1, ηZ=ηY=0.8\eta_Z=\eta_Y=0.8) without SCS or NMF. +NMF applies Nash-guided fusion to Base without CRC and SCS. These two comparisons directly evaluate the contribution of individual module to standard OT.

In addition, +SCS uses L=K=5L=K=5 and ηZ=ηY=0.8\eta_Z=\eta_Y=0.8 without Nash fusion. Since SCS selects from the candidates constructed by CRC, this variant requires CRC to provide its candidate pool. Fig. 5 further isolates the selection rule by keeping this pool and all other components fixed.

Each variant improves performance on both datasets, and the full model achieves the best results. As discussed in Model Breakdown, +NMF yields larger average gains, especially on KVC16K, supporting the contribution of Nash-guided coordination.

Dear Reviewer ddyL,

Thank you for the helpful discussion. We are glad that our responses have addressed your concerns and would welcome if any further comments or questions.

Best regards,
The Authors

Reviewer_Q7oAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NaCOT, an optimal-transport (OT) framework for multi-modal knowledge graph fusion that tackles two problems in existing OT-based methods: low target relevance (noisy, off-target text and image content) and modality imbalance (one modality dominating the fused representation). NaCOT first builds multi-scale compressed candidates of modality and fused representations, then uses sliced Wasserstein distance to select compatible source–target scales for OT alignment. Finally, it treats modalities as players in a Nash bargaining game that iteratively refines the transport plans with adaptive, inverse-utility weights.

Critical strengths
  • The motivation is clear and backed by quantitative analysis of both low target relevance and modality imbalance.
  • Applying Nash bargaining at the transport-plan level with adaptive inverse-utility weights removes the need for hand-set λ and beats a gradient-based Nash variant.
  • The experiments are comprehensive, covering ablations, robustness, hyperparameter sensitivity, backbone changes, complexity, and transfer to entity linking, with code provided.
Critical weaknesses
  • The method operates only at the sample level with a reconstruction-based utility, leaving feature-level refinement and richer utilities to future work.
  • Low target relevance is only addressed indirectly through compression and distribution matching, with no explicit filtering and no direct evidence that the selected candidates are cleaner.
  • The efficiency claim over OTKGE is not fully convincing because the SVD, CRC, and SWD selection overheads are not clearly accounted for.
Questions for authors
  • Does SCS actually select cleaner candidates? Can author provide more cases?
  • For CRC and SCS, when and how often is selection performed, and what does it select?
  • How does NaCOT behave when a modality is truly uninformative or adversarial?
  • Are the efficiency claims fair? Maybe the efficiency experiments could be further improved.
5 comments
AuthorsOpenReview ↗

[Part 1] Response to Reviewer Q7oA

W1 The method operates only at the sample level with a reconstruction-based utility, leaving feature-level refinement and richer utilities to future work.

We thank the reviewer for this observation. As acknowledged in Footnote 3 (Line 215) and the Conclusion (Lines 482--483), NaCOT leverages sample-level candidate construction and reconstruction-based utility, while feature-level refinement and alternative utility formulations are left for future refinement after we verify the viability of the framework proposed in this paper. Yet, neither is restricted by the proposed framework, and below illustrates two straightforward extensions.

(1) Feature-level refinement. A feature gating mechanism can be incorporated to refine modality representations before CRC. For instance, structural representations can be employed as a reference to co-identify target-relevant features in textual or visual inputs. The visual modality serves as an example, while the same mechanism applies to other modalities. Following Eq.(3), visual and structural features are projected into a shared embedding space: \begin{equation} H_{\mathrm{vis}} =\widetilde X_{\mathrm{vis}}W_{\mathrm{vis}} \in\mathbb R^{n_{\mathrm{vis}}\times d}. \end{equation}

\begin{equation} H_{\mathrm{str}} =\widetilde X_{\mathrm{str}}W_{\mathrm{str}} \in\mathbb R^{N\times d}. \end{equation}

Since each entity may be associated with multiple visual samples but has a single structural representation, the structural embedding is replicated and matched to all visual samples of the same entity. That is, let π(s)\pi(s) denote the entity associated with visual sample ss. The sample-aligned structural representation is constructed as Svis[s,:]=Hstr[π(s),:]S_{\mathrm{vis}}[s,:]=H_{\mathrm{str}}[\pi(s),:], (s=1,…,nvis)(s=1,\ldots,n_{\mathrm{vis}}). Then feature-wise gating integrates the two representations via: \begin{equation} G_{\mathrm{vis}} =\sigma\left( [H_{\mathrm{vis}};S_{\mathrm{vis}}; H_{\mathrm{vis}}\odot S_{\mathrm{vis}}]W_g+b_g \right). \end{equation}

\begin{equation} X_{\mathrm{vis}} =G_{\mathrm{vis}}\odot H_{\mathrm{vis}}. \end{equation} where [ ; ][\,;\,] denotes feature-wise concatenation, ⊙\odot denotes elementwise multiplication, and σ\sigma is the sigmoid function.

(2) Alternative utility formulation.

In addition to the reconstruction-based expense in Eq.(12), the Nash formulation also accommodates other convex expense functions. For instance, a contrastive similarity-based expense can be applied to encourage each transported modality representation to align more closely with its corresponding entity than with other entities:

exp⁡msim(Pm,Y)=−1N∑i=1Nlog⁡exp⁡(am,i⊤yi/τs)∑j=1Nexp⁡(am,i⊤yj/τs)+γR(Pm).\operatorname{exp}^{\mathrm{sim}}_m(P_m,Y)=-\frac{1}{N}\sum_{i=1}^{N}\log\frac{\exp(\mathbf{a}_{m,i}^{\top}\mathbf{y}_i/\tau_s)}{\sum_{j=1}^{N}\exp(\mathbf{a}_{m,i}^{\top}\mathbf{y}_j/\tau_s)}+\gamma R(P_m).

where am,i\mathbf{a}_{m,i} denotes the ii-th row of Pm⊤XmP_m^\top X_m, yi\mathbf{y}_i is the corresponding fused entity representation, and τs=0.1\tau_s=0.1 is a temperature parameter.

Notably, for fixed YY, the contrastive term is convex in PmP_m, since it consists of a log-sum-exp function of affine mappings minus a linear term. Together with the convex OT regularizer R(Pm)R(P_m), the resulting expense remains convex. Consequently, defining utility as the expense reduction in Eq.(13) preserves its concavity in PmP_m. The convex transport constraints and Nash bargaining formulation therefore remain applicable, following the analysis in Appendix E (Lines 1241--1247).

Table R1 shows the performance using the structure-guided feature refinement and the similarity-based utility function. All variants use identical data splits, backbones, fusion dimensions (d=300d=300), and training configurations. As shown, all three variants achieve performance comparable to the original NaCOT. Additionally, structure-guided feature refinement seems to provide additional performance gains, suggesting that incorporating fine-grained, entity-specific structural evidence can further enhance multimodal representation fusion. These results demonstrate the flexibility of NaCOT in accommodating finer-grained feature refinement and alternative utility formulations without substantially affecting its effectiveness.

Table R1. Individual and joint effects of structure-guided feature attention and the similarity-based utility.

MethodMKG-W MRRMKG-W Hit@1KVC16K MRRKVC16K Hit@1
NaCOT (Table 1)41.4932.1627.5517.33
NaCOT + structure-guided feature attention41.6332.2827.7717.50
NaCOT with similarity-based utility41.3832.0827.5817.34
NaCOT + structure-guided feature attention with similarity-based utility41.7032.3427.5317.29

We have included these additional experiments and the corresponding findings in Appendix G.7 of the revised manuscript.

AuthorsOpenReview ↗

[Part 2] Response to Reviewer Q7oA

W2 Low target relevance is only addressed indirectly through compression and distribution matching, with no explicit filtering and no direct evidence that the selected candidates are cleaner.

We thank the reviewer for raising this concern. Low-target relevance does not necessarily imply that a modality sample is entirely uninformative. A sample may contain substantial target-irrelevant content while retaining useful entity-specific evidence. For instance, an image dominated by background may still capture a small but informative region of the target entity, as reflected in Fig.1(b). Explicitly filtering such samples may therefore discard complementary information useful for multimodal fusion.

Motivated by this observation, NaCOT addresses low-target relevance through information-preserving candidate construction and compatibility-aware selection, rather than hard sample filtering. Specifically, CRC constructs multi-scale candidates through learnable linear combinations, allowing information from partially relevant samples to be integrated. SCS subsequently selects the source--target candidate pair with the smallest SWD, avoiding indiscriminate OT alignment across potentially incompatible representations.

Empirically, Fig.5 compares candidate selection strategies using the same candidate pool and unchanged remaining components. SWD consistently outperforms random selection and achieves performance comparable to exhaustive Greedy selection, supporting its effectiveness in identifying suitable candidates for OT fusion. Fig.4 further demonstrates the benefits of moderate compression, while Appendix G.2 shows that NaCOT maintains the highest MRR among the compared methods across all evaluated noise ratios on DB15K.

We have added a dedicated Discussion section to the revised manuscript to elaborate on the connection between candidate construction, distributional compatibility, and the influence of target-irrelevant information.

AuthorsOpenReview ↗

[Part 3] Response to Reviewer Q7oA

W3 The efficiency claim over OTKGE is not fully convincing because the SVD, CRC, and SWD selection overheads are not clearly accounted for.

We thank the reviewer for this comment. The costs of source-side construction, target-side truncated SVD, and SWD selection are discussed in Appendix F (Lines 1315–1318). Source-side construction and target-side truncated SVD are part of CRC, while SWD selection is part of SCS. For modality mm, CRC first transforms modality features into cascaded representations, which costs O(nm2d)\mathcal{O}(n_m^2d). The target-side compression is performed on YY using truncated SVD (cost O(N2d)\mathcal{O}(N^2d)).

For SWD, according to [1], with nm≥Nn_m\geq N, each candidate pair costs O(Pnmd+Pnmlog⁡nm)\mathcal{O}(Pn_md+Pn_m\log n_m), where PP is the number of projection directions. Comparing L×KL\times K candidate pairs per modality therefore gives O(PLK∑m(nmd+nmlog⁡nm))\mathcal{O}\left(PLK\sum_m(n_md+n_m\log n_m)\right). Our experiments use fixed P=100P=100 and L=K=5L=K=5, so PLKPLK is absorbed into the big-O notation. When dd dominates log⁡nm\log n_m, retaining the dominant projection term further simplifies the cost to O(∑mnmd)\mathcal{O}\left(\sum_m n_md\right), which is lower-order relative to the quadratic source-side CRC cost. The role and computational complexity of each stage of NaCOT are summarized in Table R2.

Table R2. Operations and computational complexity of each stage of NaCOT.

ModuleOperationComputational cost
CRCSource-side cascaded representation constructionO(∑mnm2d)\mathcal{O}(\sum_m n_m^2d)
CRCTarget-side compression (truncated SVD)O(∑mN2d)\mathcal{O}(\sum_m N^2d)
SCSSWD scoring and candidate selectionO(∑m(nmd))\mathcal{O}\left(\sum_m(n_md)\right)
Initial OTPairwise cost matrix constructionO(∑mnmNd)\mathcal{O}(\sum_m n_mNd)
Initial OTSinkhorn solverO(∑mTsinknmN)\mathcal{O}(\sum_m T_{\mathrm{sink}}n_mN)
NMFTransport-aligned representation and utility evaluationO(∑mnmNd)\mathcal{O}(\sum_m n_mNd)

The training times in Table 2 are measured per epoch for both NaCOT and the compared methods. NaCOT takes 74 s per epoch, compared with 76 s per epoch for OTKGE [2]. As discussed in Lines 1349–1357, this is primarily because OTKGE applies the Sinkhorn solver directly to the full-scale transfer matrix Pm∈Rnm×NP_m\in\mathbb{R}^{n_m\times N}, whereas CRC and SCS allow NaCOT to solve a smaller OT problem using selected compressed source–target pairs. With the default parameters ηZ=ηY=0.8\eta_Z=\eta_Y=0.8, even with just one compression step, the number of transport elements can be reduced to approximately 0.8×0.8=64%0.8\times0.8=64\% of the original size. Furthermore, with the default setting (T=3T=3, Appendix G.3), NMF achieves nearly saturated performance within three iterations, limiting its additional computational overhead.

We have explicitly stated the computational complexity of SWD in Appendix F of the revised manuscript. Together with the measured training times, this confirms that NaCOT maintains computational efficiency comparable to OTKGE despite its additional components.

[1] Nguyen et al. Hierarchical Sliced Wasserstein Distance. ICLR, 2023.

[2] Cao et al. OTKGE: Multi-modal Knowledge Graph Embeddings via Optimal Transport. NeurIPS, 2022.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear AC, SAC, and Reviewers:

We sincerely thank all three reviewers for their careful evaluations and positive assessments of our work. Reviewer Q7oA highlighted the clear, quantitatively supported motivation, the benefits of Nash bargaining with adaptive inverse-utility weights, and the comprehensive experiments. Reviewer ddyL recognized the importance of the two challenges, the reasonableness of selecting suitable representation scales before fusion, and the breadth of the experimental evaluation. Reviewer K68o described the work as solid and significant, emphasizing the careful analysis of the two challenges, the corresponding methodological solutions, and the accompanying theoretical analysis and experimental results. We also appreciate the reviewers' constructive concerns, which have guided our revisions.

Our main contribution is NaCOT, a Nash-guided compatible optimal transport framework that jointly addresses low-target relevance and modality imbalance in multimodal knowledge graph fusion. It combines cascaded representation construction (CRC) to generate compact, multi-scale candidates, SWD-based candidate selection (SCS) to identify distributionally compatible pairs for OT alignment, and Nash-guided modality fusion (NMF) to adaptively refine transport plans under fusion and OT constraints. We provide theoretical analysis and extensive evaluation on four MMKG benchmarks, where NaCOT achieves state-of-the-art performance, with average relative improvements of 24.68% in MRR and 23.51% in Hit@1 over the strongest baselines.

We have uploaded the revised manuscript, with the major changes highlighted in red. The main revisions are summarized below, in their order of appearance, together with the reviewer concerns they address.

  • Improving the model visualization (Fig. 2; Reviewer ddyL, W2): We have redrawn the workflow to illustrate modality projection and CRC, SWD-based candidate selection, initial OT and full-scale transport plan recovery, iterative Nash-guided modality fusion, and downstream MMKG link prediction.

  • Clarifying the full-scale cost matrix (Section 3, Lines 274–275; Reviewer ddyL, W3): We have distinguished the full-scale cost matrix from the compressed-scale matrix and explicitly defined its entries in the regularization term of Eq. (12). We have also updated the corresponding notation in Appendix E.

  • Clarifying the workflow and downstream training (Section 3, Lines 313–323; Reviewer Q7oA, Q2): We have clarified the execution order and frequency of CRC and SCS, and added the KGE scoring function and training loss in Eqs. (18)–(19) to provide a clearer description of the end-to-end NaCOT workflow.

  • Explaining candidate construction and selection (Section 5, Lines 492–526; Reviewer Q7oA, W2/Q1; Reviewer K68o, concern about CRC/SWD suitability): We have added a dedicated Discussion section explaining why partially relevant samples can still contain useful entity-specific evidence, how CRC integrates their representations, and how SCS selects compatible pairs for OT alignment. The discussion connects these choices to the existing controlled comparisons and robustness results, while distinguishing distributional compatibility from a direct measure of candidate cleanliness.

  • Updating recent literature (Appendix A; Reviewer ddyL, W4): We have expanded and reorganized the related work, incorporating 11 papers published in 2026: nine MMKG studies, one OT-based cross-modal fusion study, and one Nash bargaining-based optimization study.

  • Clarifying computational costs and timing (Appendix F, Lines 1460–1461, and Table 2; Reviewer Q7oA, W3/Q4): We have explicitly stated the computational complexity of SWD and identified it as a lower-order term relative to CRC. We have also clarified in the table caption that the reported training times are measured per epoch.

  • Formulating feature and utility extensions (Appendix G.7, Eqs. (22)–(24); Reviewer Q7oA, W1): We have added formulations for structure-guided feature refinement and a contrastive similarity-based utility, explained their integration into NaCOT, and specified a comparison design that examines them separately and jointly.

  • Correcting weight normalization (Algorithm 1, Line 15; Reviewer ddyL, W1): We have restored the normalization factor in the feasible bargaining set, making the pseudocode consistent with Eq. (15).

We are pleased that our additional experiments and clarifications have successfully addressed the main concerns of Reviewers (Q7oA) and (ddyL) , and sincerely appreciate their positive follow-up feedback and continued support.

We sincerely appreciate the reviewers’ suggestions, which have made the manuscript clearer and more precise. Detailed responses to each concern are available in the corresponding reviewer threads. We welcome any further comments or questions.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.