acceptodds
Under review as a conference paper at ICLR 2027

DSF2 :DUAL-SPACE GUIDED DISCRIMINATIVE HIGH AND LOW FREQUENCY FUSION NETWORK FOR MULTIMODAL KNOWLEDGE GRAPH COMPLETION

Abstract

Multimodal knowledge graph completion (MMKGC) aims to complete missing facts in knowledge graphs by jointly leveraging multimodal information associated with entities. Existing methods primarily fuse multimodal entity embeddings in the spatial domain, yet lack explicit modeling of frequency characteristics at different scales of variation. To capture these multiscale characteristics, shifting to the frequency domain provides a new perspective by employing spectral decomposition and filtering to facilitate multimodal fusion, but such methods neglect the pervasive coupling of spectral signals at the boundaries and the distinct characteristics of different frequencies, thereby hampering the precise characterization of multimodal features and limiting the exploitation of multimodal spectral information. To address these limitations, we propose DSF2, Dual-Space Guided Discriminative High and Low Frequency Fusion Network for MMKGC, an innovative framework that leverages different spatial geometrical properties to disentangle the pervasive inter-frequency coupling and applies distinct modeling strategies tailored to band-specific characteristics. Concretely, we design Dual-Space Frequency Disambiguation (DSFD), which better models the coupling of spectral signals by leveraging the geometric differences of corresponding spectral embeddings across these spaces to derive more discriminative features. We further design distinct processing strategies for differentiated frequencies, including a low-frequency Shared-memory Interaction Strategy (LSIS) and a High-frequency Relation-aware Gating Strategy (HRGS). Specifically, LSIS captures stable global low-frequency semantics through shared-memory interaction to promote cross-modal semantic communication and enhance multimodal consistency. HRGS focuses on high-frequency semantics characterized by local variations, employing a relation-aware gating mechanism to emphasize local information discrepancies and strengthen relation-specific high-frequency representations. This distinct strategy effectively captures differentiated frequency characteristics, thereby fully exploiting the complementary low- and high-frequency semantics. Extensive experiments on multiple benchmark datasets consistently validate the effectiveness and robustness of DSF2.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Knowledge Graph Completion, Multimodal Knowledge Graph, Frequency Fusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nKAiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DSF² for multimodal knowledge graph completion (MMKGC). The method combines entity relations, images, and textual descriptions to predict missing triples in a knowledge graph.

Critical strengths
  • The paper provides clear and well-designed visualizations that help explain the proposed method.
  • The relation-conditioned fusion design is reasonable. Different relations may require information from different modalities and at different scales, which fits the nature of the task.
  • According to the reported results, the improvements are fairly consistent. The main ranking metrics improve across all three datasets, with particularly noticeable gains on DB15K.
Critical weaknesses
  • It is unclear how useful multimodal knowledge graph completion remains in real-world applications today. The paper does not provide enough case studies to demonstrate its practical value.
  • The authors should carefully check whether the released code is consistent with the model described in the paper. For example, it is worth verifying whether the implementation of DSFD actually includes the proposed energy grouping, hyperbolic space mapping, and dual-space contrastive learning components.
  • On pages 7–8, SNAG is categorized as an "LLM-based MMKGC" method, while on page 13, it is grouped with generative or hybrid completion methods. However, the original SNAG is a representation learning framework that uses modality noise, Transformer-based interactions, and RotatE for completion scoring. These classifications do not seem accurate.
  • The discussion and comparison with existing work on hyperbolic representations and multimodal knowledge graph completion are not sufficiently comprehensive.
  • The paper seems to be largely incremental, building on relatively conventional techniques.
  • The performance improvements are relatively limited.
Questions for authors
  • Given the rapid development of multimodal large language models (MLLMs), can these models also perform multimodal knowledge graph completion? What advantages might they offer compared with the proposed method?
1 comment

Reply to W1:Thank you for raising this concern. We agree that benchmark link-prediction results alone do not sufficiently illustrate the practical value of multimodal knowledge graph completion. In e-commerce, MPKGAC formulates missing product attributes as a multimodal KG completion task and reports a 2.14% relative increase in click-through rate in an Alibaba application (Wang et al., 2023). In biomedicine, Xiao et al. (2026) study the completion of multimodal biomedical KGs, where missing relations are relevant to scientific investigation. In cultural heritage, ArtKGC examines how completed artwork facts can provide evidence for artwork question answering and explanation (Zhou et al., 2026). These studies demonstrate the value of recovering missing facts when structured knowledge is incomplete but complementary evidence is available.

Reply to W2:Thank you for raising this concern. We agree that consistency between the released code and the described model should be verified explicitly. The energy-based grouping is examined in Appendices C and L, the hyperbolic mapping is evaluated in Section 4.5 and Appendix H, with additional before-and-after DSFD analysis in Appendix I, and the dual-space objective is examined in the ablation study and Appendix H. To address the implementation concern directly, we will also check the released code against the corresponding equations and provide a clear mapping to the functions implementing energy grouping, Poincaré mapping, and the Euclidean and hyperbolic loss terms.

Reply to W3:We thank the reviewer for pointing out this misclassification. We agree that SNAG uses pretrained models for feature extraction but is not an LLM-based MMKGC method. We have corrected its classification in the appendix related work and replaced SNAG with LBMKGC in the main comparison table.

Reply to W4:We thank the reviewer for highlighting this gap. We have expanded the discussion of prior work on hyperbolic representations and MMKGC, and added further analyses in Appendices H, I, J, and K. These additions clarify the motivation for our geometric design and provide additional empirical evidence for its behavior.

Reply to W5:First, very few studies in the field of knowledge graphs have taken a frequency-domain perspective. In our review of the past three years, we found only about four or five such papers, and WFF is the only one published at a top-tier conference. Moreover, WFF directly transfers spatial-domain processing logic to the frequency domain without treating low- and high-frequency components separately according to their distinct characteristics.

Reply to W6:Frequency-domain methods show a clear advantage on DB15K, whose richer relation types and denser graph connectivity provide informative structural patterns for modeling smooth shared semantics and relation-specific local variations, which semantic enhancement relying mainly on large pretrained models struggles to fully capture. LLM-based methods achieve several best or second- best results on MKG-W and MKG-Y, but their performance drops on DB15K, with some metrics below those of conventional multimodal KGC methods. DSF2 outperforms most LLM-based results on MKG-W and MKG-Y and improves over the frequency-domain baseline on DB15K, demonstrating its ability to model both shared smooth semantics and local discriminative variations.

Reply to Q1:Yes, MLLMs can perform multimodal knowledge graph completion. HFR-MKGC, which is included in our main comparison, uses LLaVA to process visual information and incorporates the reasoning output of a fine-tuned MLLM into completion scoring. MLLMs can draw on broad pretrained visual–language knowledge and support flexible, triple-specific reasoning. HFR-MKGC performs competitively on MKG-W and MKG-Y, but its results on DB15K are weaker, with some metrics below conventional MMKGC methods. In Table 1, DSF² achieves higher MRR, Hits@1, and Hits@3 than the LLM-based comparators on MKG-W and MKG-Y, and outperforms the frequency-domain baselines on DB15K. These results support the effectiveness of explicitly modeling shared low-frequency semantics and relation-discriminative high-frequency variations, while recognizing the complementary strengths of MLLM-based reasoning.In future work, we will also explore how multimodal large language models can be integrated into our framework.

Reviewer_4sMkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DSF², which is a frequency-domain framework for multimodal knowledge graph completion (MMKGC) that integrates graph spectral decomposition with hyperbolic geometry to better handle frequency-ambiguous samples. Specifically, it employs graph wavelets to decompose multimodal entity representations into low- and high-frequency bands and introduces a band-energy-based routing mechanism that maps ambiguous samples into hyperbolic space, where a cross-space contrastive loss enhances their discriminability. The resulting representations are further refined through a shared memory bank and relation-gated scale selection. By combining frequency-aware sample routing with dual-space representation learning, DSF² provides an explicit mechanism for modeling instances without clear frequency dominance. Extensive experiments on MKG-W, MKG-Y, and DB15K demonstrate that DSF² outperforms 13 baselines in MRR, Hits@1, and Hits@3, while ablation studies further validate the effectiveness of its dual-space routing, LSIS, and HRGS components.

Critical strengths
  1. Scoring each instance by band energy and moving only ambiguous ones into hyperbolic space goes beyond the uniform processing of existing frequency-domain methods and is a new idea in this line of work.
  2. The method achieves the best MRR, Hits@1, and Hits@3 on all three benchmarks against 13 baselines, and the results hold up across datasets with ablations covering every component.
Critical weaknesses
  1. The transition band is defined by the overlap of the low-pass and band-pass kernels, and q is computed from their outputs. Whether an entity–modality instance is selected as ambiguous therefore depends on the kernels' shape and scales as well as on its spectral energy distribution. The paper builds dedicated routing and hyperbolic modules around these instances, but does not discuss how much of the ambiguity comes from the decomposition itself. If a large part stems from the filter design, DSFD is partly addressing a problem the method creates.
  2. The roles assigned to the hyperbolic mapping and the memory banks are not backed by the design. The paper describes Poincaré geometry as providing "a flexible non-Euclidean metric structure", but does not say which property of this geometry suits instances without a clear frequency dominance. The three LSIS memory banks share the same structure, inputs, and update formulas, and no bank-specific constraint is imposed, so nothing ties each bank to the role its name implies.
Questions for authors
  1. What fraction of entity–modality instances falls into B_A on each dataset and modality? How much does this set change when the kernel scales s_j or the low-pass kernel are varied? It would also help to know how B_A differs from B_L and B_H in degree distribution, entity and relation types, and agreement between modalities. These points bear directly on whether the ambiguity DSFD targets is a property of the data or of the chosen filter bank. They also determine how much of the reported gain the dual-space module can account for.
  2. Which property of the Poincaré ball do the authors consider responsible for the improvement on ambiguous instances? Is there any evidence that these instances have the kind of structure hyperbolic space is usually suited to? Table 3 shows that the mapping helps. A clear explanation, however, would determine whether the dual-space design is a principled contribution or an empirical choice.
  3. The three banks share the same architecture, inputs, and update rule. Is there any analysis showing that M^g, M^r, and M^tr store the global, relation-related, and cross-modal information their names describe? In particular, are the routing weights in Figure 9 measured from the trained model, or constructed for illustration, as the caption ("using the specified values") suggests?
  4. RASD and MMQuat are also built on spectral decomposition but are only cited in Appendix A.2. Including them in the main comparison, even on a subset of the datasets, would strengthen the claims of superiority in this line of work.
  5. The final loss consists of two parts: the base KGC loss and the DSFD loss. I suggest giving the explicit formulas for both parts in the main text.
2 comments

Reply to W1&Q1:We thank the reviewer for this question. We added a dataset- and modality-wise analysis of the frequency-polarity groups in the Appendix L. Each available entity–modality pair is counted once, and the corresponding figure shows the relative shares assigned to BL\mathcal{B}_L, BA\mathcal{B}_A, and BH\mathcal{B}_H. This identifies the scope of the ambiguous branch in DSFD.

Reply to W2:The Poincaré ball has a position-dependent metric, its local distance scale increases toward the boundary. In DSFD, clear low- and high-frequency instances are mapped into the ball as references, and hyperbolic distances constrain ambiguous instances relative to positive and negative references. This provides a geometric inductive bias for distinguishing embeddings that may be close in Euclidean space. Poincaré geometry provides a nonlinear geodesic metric for comparing ambiguous instances with clear-frequency references, as visualized in the new Appendix H.

Reply to W3&Q2:We agree that routing weights alone cannot establish the semantic contents of the three banks. We therefore plan a bank-level content and intervention analysis across multiple seeds. For MgM^g, we will record the relation and modality coverage of its writes, test the stability of its readouts when the query relation changes, and examine whether masking its readout has broadly distributed effects. For MrM^r, we will vary the relation for the same entity–modality pair while holding other inputs fixed, and test whether its readout and the effect of masking it are particularly relation-sensitive. For MtrM^{tr}, we will start from identical memory snapshots, write from one modality, and measure subsequent retrieval and prediction through another modality; writes from an unrelated entity will serve as a control.The phrase “using the specified values” was an editing error that we failed to catch during proofreading. We apologize for the confusion. The routing weights in Figure 9 were computed from the trained model, rather than constructed for illustration.

Reply to W4:We thank the reviewer for this valuable suggestion. We agree that including RASD and MMQuat would strengthen the comparison with spectral methods. RASD reports results on structure-only KGC benchmarks, which cannot be directly transferred to our multimodal benchmarks. For MMQuat, we have not yet established a matched evaluation under our data splits, modality features, and ranking protocol. We plan to reproduce or adapt both methods on common benchmarks, clearly identify RASD as a structure-only spectral baseline, and add the resulting comparisons to the main table in a subsequent revision.

Reply to W5:Thank you for the suggestion. The DSFD loss was already explicitly defined in Eq. (4) of the original manuscript. We have now added the explicit base KGC loss in Eq. (12) and clarified how the two losses form the final training objective.

Thanks for the response. I have reviewed Appendix L. The results show that frequency ambiguity is common across datasets and modalities. However, it remains unclear whether this ambiguity comes from the data itself or the filter design. Further experiments are needed to examine how the distributions of clear and ambiguous instances change with the wavelet scale parameters (s_j). Regarding RQ3 and RQ4, the authors have only outlined their plans for future experiments. No substantial revisions have been made yet. Moreover, experimental validation alone is insufficient to address RQ3. A more principled design should introduce bank-specific inputs, read/write mechanisms, or training constraints based on the intended function of each memory bank. Therefore, I maintain my original score.

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

This paper proposes DSF² via integrating Dual-Space Frequency Disambiguation (DSFD), Low-frequency Shared-memory Interaction (LSIS), and High-frequency Relation-aware Gating (HRGS).

Critical strengths

The paper provides an interesting perspective on frequency-aware multimodal fusion by separately modeling low- and high-frequency characteristics, supported by comprehensive experiments.

Critical weaknesses
  1. The method involves numerous trainable projection matrices, cross-space contrastive learning and hierarchical memory interactions. However, the paper lacks a detailed analysis of model parameters, time complexity, memory consumption, and training/inference costs. It is unclear whether the improvements primarily arise from the proposed method or simply from increased model capacity.

  2. In Section 3.3 and Equation (4), clear-frequency instances are modeled in Euclidean space, whereas boundary-ambiguous instances are mapped into Poincaré space. Why is hyperbolic geometry particularly suitable for frequency ambiguity? Although Table 3 demonstrates that this configuration performs better than others, these results do not fully explain the underlying rational. Moreover, the frequency polarity in Equation (3) measures relative low- and high-frequency energy dominance rather than spectral coupling directly. Can authors explain why such definition works? Can authors establish the theoretical relationship between proposed frequency polarity and spectral coupling?

  3. Appendix C shows that DSFD brings larger performance gains to samples with stronger transition responses. However, these observations do not directly demonstrate that the reconstructed representations exhibit reduced coupling between low- and high-frequency components.

  4. The method assumes that low-frequency components capture stable, shared cross-modal semantics, while high-frequency components capture local variations and relation-discriminative information. Please provide the reason or if any existing work for supporting. Furthermore, does each modality use a different graph Laplacian? What happen to the comparability of frequency characteristics across modalities?

  5. Some important ablations are missing. For example, λDSFD\lambda_{\mathrm{DSFD}}. Similarly, the effects of the Euclidean and contrastive temperatures and the hyperbolic curvature parameter remain unclear.

  6. Previous approaches such as WFF already employ wavelet-based frequency modeling, while MCKGC explores multimodal representation learning across different geometric spaces. The proposed DSFD combines related ideas, but, again, it remains unclear whether its improvements arise from a fundamentally different frequency-disambiguation mechanism or from additional contrastive constraints and nonlinear transformations.

  7. LSIS maintains shared memory states across training mini-batches, which may introduce sensitivity to mini-batch ordering and memory initialization. Can authors examine the stability of this mechanism under different update settings?

Questions for authors

see Weakness

1 comment

Reply to W1:Thank you for raising the computational-cost and model-capacity concern. We have added an analysis in Appendix G comparing DSF² with MoMoK and a parameter-matched variant, MoMoK-PM, on all three datasets. We distinguish the one-time spectral preprocessing from recurring training computations and report trainable parameters, per-epoch training time, full-candidate inference throughput, and peak training/inference GPU memory. MoMoK-PM differs from DSF² by only 0.1M parameters. Relative to MoMoK, DSF² increases per-epoch training time by 24.0–27.4% and peak GPU memory by 3.4 GiB during training and 3.2 GiB during inference, while reducing inference throughput by 17.2–17.7%. These results make the computational and memory costs of DSF² explicit and provide a matched parameter budget for examining the capacity explanation.

Reply to w2:Equation (3) measures relative low- and high-frequency energy dominance, not spectral coupling directly .Appendix C shows an empirical association with transition responses. Poincaré geometry provides a nonlinear geodesic metric for comparing ambiguous instances with clear-frequency references, as visualized in the new Appendix H. And the energies defined in Eq. (2) are simply the squared l2 norms of the low- and high-frequency responses, “energy” is the term we use for these quantities.

Reply to W3:We agree that the performance gains in Appendix C alone do not demonstrate reduced coupling between the reconstructed branches. To address this, Appendix I introduces a paired before-and-after DSFD analysis of all frequency-ambiguous instances on MKG-W. We use a joint PCA to compare their distributions and measure low–high transition-band alignment in the original feature space. This experiment directly tests whether DSFD reduces alignment between the reconstructed branches, complementing the performance evidence in Appendix C.

Reply to W4:Frequency-domain modeling has been widely explored in computer vision, including image fusion and recognition, but remains relatively underexplored in multimodal knowledge graph completion. Studies such as CDDFuse (CVPR 2023), WaveMamba (ICCV 2025), DLVP-CLIP (CVPR 2026), FRAMER (CVPR 2026), DyFCLT (CVPR 2026), and Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition (ICCV 2025) motivate the distinction: low-frequency components often preserve global structure and cross-modal commonalities, whereas high-frequency components capture local details and potentially discriminative variations.Our spectral graph wavelet decomposition follows Hammond et al. (2011). The structural, visual, and textual modalities use distinct graph Laplacians, Ls\mathbf{L}_s, Lv\mathbf{L}_v, and Lt\mathbf{L}_t, respectively, each computed as Lm=I−Dm−1/2AmDm−1/2\mathbf{L}_m=\mathbf{I}-\mathbf{D}_m^{-1/2}\mathbf{A}_m\mathbf{D}_m^{-1/2}. They share the same entity set, normalization scheme, wavelet kernels, and scale settings, but have different adjacency structures and spectral bases Um\mathbf{U}_m.

Reply to W5:Thank you for pointing out the missing analysis. We have added a four-parameter sensitivity study in the appendix, varying the DSFD loss weight, the Euclidean and Poincaré temperatures, and hyperbolic curvature while holding other settings fixed.

Reply to W6:Although WFF separates low- and high-frequency components, it applies essentially the same multimodal processing strategy to both, without tailoring their treatment to their distinct characteristics. In Section 4.4, we test this distinction by applying a unified processing strategy to the two branches. The resulting performance degradation supports the contribution of our differentiated frequency processing mechanism.

Reply to W7:We thank the reviewer for raising this concern. We added a ten-seed stability evaluation in Appendix K, varying mini-batch order, memory initialization, update rate, and retention across epochs. Orders A–D process identical mini-batches and negative samples in different reproducible sequences. We also include a No Memory control that disables LSIS reads and writes. This comparison tests both the sensitivity of LSIS to training conditions and the contribution of persistent memory updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.