Complementarity-guided Modality Knowledge Fusion

• The paper introduces CMKF, which leverages information-theoretic complementarity to fuse multimodal embeddings for enhanced knowledge graph completion.
• It decomposes each modality into multiple expert views and uses mutual information to weight them, reducing redundancy and emphasizing unique contributions.
• The two-stage fusion process, integrating intra- and inter-modal complementarity, significantly outperforms traditional concatenation, gating, or element-wise methods.

to=arxiv_search qq天天中彩票 天天啪 {"^^^^^^^^2query2 OR \2^^^^^^^^"Complementarity-driven Representation Learning for Multi-modal Knowledge Graph Completion\"","max_results":5,"sort_by":"relevance"}【อ่านข้อความเต็มjson {"results":[{"arxiv_id":"(&&&^^^^2query2^^^^&&&)","title":"Complementarity-driven Representation Learning for Multi-modal Knowledge Graph Completion","authors":["Tong Yang","Jingfeng Wang","Shuo Han","Yutong Wang","Hongming Liu","Zekun Li","Huan Zhao","Wendi Duan"],"abstract":"Multi-modal Knowledge Graph Completion (MMKGC) aims to uncover hidden world knowledge in multimodal knowledge graphs by leveraging both multimodal and structural entity information. However, the inherent imbalance in multimodal knowledge graphs, where modality distributions vary across entities, poses challenges in utilizing additional modality data for robust entity representation. Existing MMKGC methods typically rely on attention or gate-based fusion mechanisms but overlook complementarity contained in multi-modal data. In this paper, we propose a novel framework named Mixture of Complementary Modality Experts (MoCME), which consists of a Complementarity-guided Modality Knowledge Fusion (CMKF) module and an Entropy-guided Negative Sampling (EGNS) mechanism. The CMKF module exploits both intra-modal and inter-modal complementarity to fuse multi-view and multi-modal embeddings, enhancing representations of entities. Additionally, we introduce an Entropy-guided Negative Sampling mechanism to dynamically prioritize informative and uncertain negative samples to enhance training effectiveness and model robustness. Extensive experiments on five benchmark datasets demonstrate that our MoCME achieves state-of-the-art performance, surpassing existing approaches.","categories":["cs.AI","cs.IR"],"published":"2^^^^2query22^^^^5-^^^^2query2 OR \2^^^^"}]} to=arxiv_search 天天中奖彩票 ադարձ to=arxiv_search 微信天天彩票 json {"^^^^^^^^2query2 to Combine Modalities in Multimodal Deep Learning\" OR (Liu et al., 2018)) OR (\"Fusion-Mamba for Cross-modality Object Detection\" OR (Dong et al., 2024)) OR (\"CHARM: Collaborative Harmonization across Arbitrary Modalities for Modality-agnostic Semantic Segmentation\" OR (Wen et al., 5 Aug 2025))","max_results":^^^^2id:(Li, 28 Jul 2025) OR \2query2^^^^,"sort_by":"relevance"}【อ่านข้อความเต็มjson {"results":[{"arxiv_id":"(Liu et al., 2018)","title":"Learn to Combine Modalities in Multimodal Deep Learning","authors":["John Arevalo","Thamar Solorio","Manuel Montes-y-Gómez","Fabio A. González"],"abstract":"Combining complementary information from multiple modalities is intuitively appealing for improving the performance of learning-based approaches. However, it is challenging to fully leverage different modalities due to practical challenges such as varying levels of noise and conflicts between modalities. Existing methods do not adopt a joint approach to capturing synergies between the modalities while simultaneously filtering noise and resolving conflicts on a per sample basis. In this work we propose a novel deep neural network based technique that multiplicatively combines information from different source modalities. Thus the model training process automatically focuses on information from more reliable modalities while reducing emphasis on the less reliable modalities. Furthermore, we propose an extension that multiplicatively combines not only the single-source modalities, but a set of mixtured source modalities to better capture cross-modal signal correlations. We demonstrate the effectiveness of our proposed technique by presenting empirical results on three multimodal classification tasks from different domains. The results show consistent accuracy improvements on all three tasks.","categories":["cs.LG","stat.ML"],"published":"2^^^^2query2id:(Li, 28 Jul 2025) OR \2^^^^8-^^^^2query2^^^^5-29","pdf_url":"http://arxiv.org/pdf/([1805.11730](/papers/1805.11730))v^^^^2id:([2507.20620](/papers/2507.20620)) OR \2^^^^"},{"arxiv_id":"(Dong et al., 2024)","title":"Fusion-Mamba for Cross-modality Object Detection","authors":["Chengxin Wang","Yingzheng Qin","Yueliang He","Jin Xu","Junpeng Shi","Xiaodong Gong","Yang Xiao"],"abstract":"Cross-modality fusing complementary information from different modalities effectively improves object detection performance, making it more useful and robust for a wider range of applications. Existing fusion strategies combine different types of images or merge different backbone features through elaborated neural network modules. However, these methods neglect that modality disparities affect cross-modality fusion performance, as different modalities with different camera focal lengths, placements, and angles are hardly fused. In this paper, we investigate cross-modality fusion by associating cross-modal features in a hidden state space based on an improved Mamba with a gating mechanism. We design a Fusion-Mamba block (FMB) to map cross-modal features into a hidden state space for interaction, thereby reducing disparities between cross-modal features and enhancing the representation consistency of fused features. FMB contains two modules: the State Space Channel Swapping (SSCS) module facilitates shallow feature fusion, and the Dual State Space Fusion (DSSF) enables deep fusion in a hidden state space. Through extensive experiments on public datasets, our proposed approach outperforms the state-of-the-art methods on PRESERVED_PLACEHOLDER^^^^2query2^^^^AP with 5.9% on PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^ and 4.9% on FLIR-Aligned datasets, demonstrating superior object detection performance. To the best of our knowledge, this is the first work to explore the potential of Mamba for cross-modal fusion and establish a new baseline for cross-modality object detection.","categories":["cs.CV"],"published":"2^^^^2query22^^^^4-^^^^2query2 OR \2^^^^4","pdf_url":"http://arxiv.org/pdf/([2404.09146](/papers/2404.09146))v2"},{"arxiv_id":"([2508.03060](/papers/2508.03060))","title":"CHARM: Collaborative Harmonization across Arbitrary Modalities for Modality-agnostic Semantic Segmentation","authors":["Jinxuan Yang","Leixin Yang","Peiming Sun","Ningchen Wang","Yanhong Yuan","Qinglong Zhang","Rui Huang","Xiaoqi Zhao","Kaizhu Huang"],"abstract":"Modality-agnostic Semantic Segmentation (MaSS) aims to achieve robust scene understanding across arbitrary combinations of input modality. Existing methods typically rely on explicit feature alignment to achieve modal homogenization, which dilutes the distinctive strengths of each modality and destroys their inherent complementarity. To achieve cooperative harmonization rather than homogenization, we propose CHARM, a novel complementary learning framework designed to implicitly align content while preserving modality-specific advantages through two components: (^^^^2id:(Li, 28 Jul 2025) OR \2^^^^) Mutual Perception Unit (MPU), enabling implicit alignment through window-based cross-modal interaction, where modalities serve as both queries and contexts for each other to discover modality-interactive correspondences; (2) A dual-path optimization strategy that decouples training into Collaborative Learning Strategy (CoL) for complementary fusion learning and Individual Enhancement Strategy (InE) for protected modality-specific optimization. Experiments across multiple datasets and backbones indicate that CHARM consistently outperform the baselines, with significant increment on the fragile modalities. This work shifts the focus from model homogenization to harmonization, enabling cross-modal complementarity for true harmony in diversity.","categories":["cs.CV"],"published":"2^^^^2query22^^^^5-^^^^2query2 OR \2^^^^"}]} to=arxiv_search ացին json {"^^^^^^^^2query2 Diffusion Policy: Robust Generalization via Complementarity-Aware Multimodal Fusion\" OR (&&&^^^^2id:(Li, 28 Jul 2025) OR \22^^^^&&&)) OR (\"IMKD: Intensity-Aware Multi-Level Knowledge Distillation for Camera-Radar Fusion\" OR (&&&^^^^2id:(Li, 28 Jul 2025) OR \2^^^^3&&&))","max_results":^^^^2id:(Li, 28 Jul 2025) OR \2query2^^^^,"sort_by":"relevance"}【อ่านข้อความเต็มjson {"results":[{"arxiv_id":"(&&&^^^^2id:(Li, 28 Jul 2025) OR \22^^^^&&&)","title":"Visual-Geometry Diffusion Policy: Robust Generalization via Complementarity-Aware Multimodal Fusion","authors":["Tianyu Wang","Mingcong Lu","Jiaming Liu","Ping Li","Yunfan Jiang","Wenxuan Jia","Xiangyu Kong","Qinyang Li","Xingxu Zhou"],"abstract":"Imitation learning has emerged as a crucial ap proach for acquiring visuomotor skills from demonstrations, where designing effective observation encoders is essential for policy generalization. However, existing methods often struggle to generalize under spatial and visual randomizations, instead tending to overfit. To address this challenge, we propose Visual Geometry Diffusion Policy (VGDP), a multimodal imitation learning framework built around a Complementarity-Aware Fusion Module where modality-wise dropout enforces balanced use of RGB and point-cloud cues, with cross-attention serving only as a lightweight interaction layer. Our experiments show that the expressiveness of the fused latent space is largely induced by the enforced complementarity from modality-wise dropout, with cross-attention serving primarily as a lightweight interaction mechanism rather than the main source of robustness. Across a benchmark of ^^^^2id:(Li, 28 Jul 2025) OR \2^^^^8 simulated tasks and 4 real-world tasks, VGDP outperforms seven baseline policies with an average performance improvement of 39.^^^^2id:(Li, 28 Jul 2025) OR \2^^^^%. More importantly, VGDP demonstrates strong robustness under visual and spatial per turbations, surpassing baselines with an average improvement of 4^^^^2id:(Li, 28 Jul 2025) OR \2^^^^.5% in different visual conditions and ^^^^2id:(Li, 28 Jul 2025) OR \2^^^^5.2% in different spatial settings.","categories":["cs.RO","cs.AI"],"published":"2^^^^2query22^^^^5-^^^^2id:(Li, 28 Jul 2025) OR \2id:(Li, 28 Jul 2025) OR \2^^^^-27","pdf_url":"http://arxiv.org/pdf/(&&&^^^^2id:([2507.20620](/papers/2507.20620)) OR \22^^^^&&&)v^^^^2id:(Li, 28 Jul 2025) OR \2^^^^"},{"arxiv_id":"(&&&^^^^2id:(Li, 28 Jul 2025) OR \2^^^^3&&&)","title":"IMKD: Intensity-Aware Multi-Level Knowledge Distillation for Camera-Radar Fusion","authors":["Muhammad Adnan","Maksym Klymenko","Marrakchi Yassine","Nico M. Schmidt","Patrick Mäder"],"abstract":"High-performance Radar-Camera 3D object detection can be achieved by leveraging knowledge distillation without using LiDAR at inference time. However, existing distillation methods typically transfer modality-specific features directly to each sensor, which can distort their unique characteristics and degrade their individual strengths. To address this, we introduce IMKD, a radar-camera fusion framework based on multi-level knowledge distillation that preserves each sensor's intrinsic characteristics while amplifying their complementary strengths. IMKD applies a three-stage, intensity-aware distillation strategy to enrich the fused representation across the architecture: (^^^^2id:(Li, 28 Jul 2025) OR \2^^^^) LiDAR-to-Radar intensity-aware feature distillation to enhance radar representations with fine-grained structural cues, (2) LiDAR-to-Fused feature intensity-guided distillation to selectively highlight useful geometry and depth information at the fusion level, fostering complementarity between the modalities rather than forcing them to align, and (3) Camera-Radar intensity-guided fusion mechanism that facilitates effective feature alignment and calibration. Extensive experiments on the nuScenes benchmark show that IMKD reaches 67.^^^^2query2^^^^% NDS and 6^^^^2id:(Li, 28 Jul 2025) OR \2^^^^.^^^^2query2^^^^% mAP, outperforming all prior distillation-based radar-camera fusion methods. Our code and models are available at https://github.com/dfki-av/IMKD/.","categories":["cs.CV"],"published":"2^^^^2query22^^^^5-^^^^2id:([2507.20620](/papers/2507.20620)) OR \22^^^^-^^^^2id:(Li, 28 Jul 2025) OR \2^^^^7","pdf_url":"http://arxiv.org/pdf/(&&&^^^^2id:([2507.20620](/papers/2507.20620)) OR \2^^^^3&&&)v^^^^2id:(Li, 28 Jul 2025) OR \2^^^^"}]} Complementarity-guided Modality Knowledge Fusion (CMKF) is the representation-learning module introduced by Tong Yang, Jingfeng Wang, Shuo Han, Yutong Wang, Hongming Liu, Zekun Li, Huan Zhao, and Wendi Duan in the MoCME framework for multi-modal knowledge graph completion (MMKGC). In that formulation, CMKF addresses a specific limitation of prior MMKGC systems: existing methods typically rely on attention or gate-based fusion mechanisms but overlook complementarity contained in multi-modal data. CMKF therefore treats fusion as a complementarity-aware mixture-of-experts process, modeling both intra-modal complementarity among expert views within a modality and inter-modal complementarity across modalities, and producing a joint entity embedding that is used by a RotatE-based score function for link prediction (&&&^^^^2query2^^^^&&&).

^^^^2id:(Li, 28 Jul 2025) OR \2^^^^. Problem setting and conceptual motivation

MMKGC aims to uncover hidden world knowledge in multimodal knowledge graphs by leveraging both multimodal and structural entity information (&&&^^^^2query2^^^^&&&). The multimodal knowledge graph is defined as

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$

where $\mathcal{E}$ denotes entities, $\mathcal{R}$ relations, $\mathcal{T}=\{(h,r,t)\}$ triples, and $\mathcal{M}$ modalities. For an entity $e \in \mathcal{E}$ and modality $m \in \mathcal{M}$, the raw modality data are

$$\mathcal{I}_m(e)=\{\mathcal{I}_{e,m}^{(1)}, \mathcal{I}_{e,m}^{(2)}, \dots\}.$$

The paper’s central motivation is modality imbalance. Different entities have different modality coverage: some entities may have images and text, others may only have one modality, and some modalities may be missing, noisy, or incomplete (&&&^^^^2query2^^^^&&&). In that setting, simple concatenation, gating, or attention is presented as insufficient. Concatenation mixes everything without measuring usefulness or redundancy; gating learns importance weights but does not explicitly represent complementarity; and attention can be expensive in large MMKGs, especially when pairwise interactions must be computed at scale (&&&^^^^2query2^^^^&&&).

CMKF is built around a more specific question than global modality importance. The module asks which views within a modality are least redundant with the others, and which modalities are most complementary to the rest. This is the defining sense in which CMKF is “complementarity-guided”: it operationalizes uniqueness and redundancy rather than relying on a black-box weighting mechanism (&&&^^^^2query2^^^^&&&).

2. Position inside MoCME and modality encoding pipeline

MoCME consists of two major parts: CMKF, which constructs robust entity representations from multimodal inputs, and Entropy-guided Negative Sampling (EGNS), which improves training. CMKF is the representation learning backbone. It produces a final joint entity embedding used later by the KG completion scoring function, specifically RotatE (&&&^^^^2query2^^^^&&&).

Before fusion, each modality PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2query2^^^^ is encoded by a pretrained encoder PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2id:(Li, 28 Jul 2025) OR \2^^^^: PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \22^^^^ The raw modality feature is then projected into a common embedding space by a modality-specific projection network PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^3: PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^4 The implementation details given in the paper are specific: the visual modality uses VGG^^^^2id:(Li, 28 Jul 2025) OR \2^^^^6, the textual modality uses BERT, the encoders are frozen, and the projection uses a two-layer MLP with ReLU. For the structural modality, the final structural embedding PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^5 is initialized as a learnable parameter rather than being frozen (&&&^^^^2query2^^^^&&&). The paper also treats the structural KG information as another modality.

This stage standardizes heterogeneous sources in a shared dimensional space before expert decomposition and complementarity fusion. CMKF then transforms raw modality inputs into modality-specific embeddings, multi-view expert embeddings within each modality, an intra-modality fused embedding per modality, and finally a joint multimodal embedding across modalities (&&&^^^^2query2^^^^&&&).

3. Intra-modal complementarity and multi-view expert decomposition

Within a single modality, the paper argues that one embedding may not be enough to capture all semantics. CMKF therefore adopts a mixture-of-experts design with PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^6 experts per modality,

PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^7

each mapping the modality representation into a different semantic subspace and producing multi-view embeddings

PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^8

These views are intended to be diverse, so that each expert captures a different aspect of the modality (&&&^^^^2query2^^^^&&&).

The mechanism for deciding how strongly each expert view should contribute is explicitly information-theoretic. Redundancy is measured through mutual information: PRESERVED_PLACEHOLDER_^^^^2id:(Li, 28 Jul 2025) OR \2^^^^9 where

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$^^^^2query2^^^^

High mutual information is interpreted as similarity or redundancy; low mutual information is interpreted as different information and therefore greater complementarity (&&&^^^^2query2^^^^&&&).

The intra-modal fused embedding is then defined as

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$^^^^2id:(Li, 28 Jul 2025) OR \2^^^^

with normalized weights

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$2

A view therefore receives a higher weight when it has lower mutual information with the other expert views. The paper interprets this as a direct operationalization of intra-modal complementarity: views that are less redundant are more valuable, views that bring unique information are emphasized, and overly similar views are down-weighted (&&&^^^^2query2^^^^&&&). A common misunderstanding is to treat this stage as merely another gating layer. In the paper’s formulation, the weights are not described as arbitrary learned coefficients; they are derived from mutual-information-based redundancy estimates.

4. Inter-modal complementarity and joint entity embedding

After intra-modal fusion, CMKF has one embedding per modality, $$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$3 for $$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$4. It then performs a second-stage fusion across modalities. This second stage is motivated by the claim that different modalities often encode different kinds of knowledge: text may provide semantic labels or descriptions, image may capture visual appearance, audio or video may capture temporal behavior, numeric attributes may represent scalar facts, and structure captures relational semantics (&&&^^^^2query2^^^^&&&).

As in the intra-modal stage, the contribution of each modality is weighted by its mutual information with the others. The joint embedding is

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$5

where

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$6

The paper interprets this stage as selecting modalities that contribute the most non-redundant information. A modality that is very similar to the others is down-weighted, whereas a modality that provides unique signal is up-weighted; missing or noisy modalities are said to be naturally compensated for because fusion can rely on more informative ones (&&&^^^^2query2^^^^&&&). This is the paper’s answer to modality imbalance.

The final joint embedding is used by a RotatE-based KG completion score: $$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$7 where $$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$8 denotes Hadamard product in complex space (&&&^^^^2query2^^^^&&&). The paper states that RotatE is used because it models relation patterns well in complex space, including symmetry, antisymmetry, inversion, and composition. CMKF thus provides the entity representations, while RotatE scores triple plausibility.

5. Optimization interface with EGNS

Although EGNS is a separate module, the full optimization pipeline links it directly to CMKF. After CMKF produces the joint embeddings for positive and negative triples, the model computes negative-triple scores, converts them to probabilities by

$$\mathcal{MG} = (\mathcal{E}, \mathcal{R}, \mathcal{T}, \mathcal{M}),$$9

and estimates entropy as

$\mathcal{E}$^^^^2query2^^^^

Negative samples are partitioned by thresholds $\mathcal{E}$^^^^2id:(Li, 28 Jul 2025) OR \2^^^^ into easy, ambiguous, and hard categories, with type-dependent weights satisfying

$\mathcal{E}$2

The final loss is

$\mathcal{E}$3

In this architecture, CMKF affects the final loss by producing the embeddings used in $\mathcal{E}$4 and $\mathcal{E}$5 (&&&^^^^2query2^^^^&&&). CMKF is therefore not the whole MoCME framework. A frequent simplification is to equate the two, but the paper clearly separates representation fusion from entropy-guided negative sampling.

6. Empirical contribution and relation to broader complementarity-guided fusion

The experimental evidence reported for CMKF centers on component ablation and fusion comparison. On DB^^^^2id:(Li, 28 Jul 2025) OR \2^^^^5K, the full MoCME model achieves MRR 39.62, Hit@^^^^2id:(Li, 28 Jul 2025) OR \2^^^^ 29.7^^^^2id:(Li, 28 Jul 2025) OR \2^^^^, and Hit@^^^^2id:(Li, 28 Jul 2025) OR \2query2^^^^ 55.36. When the intra-modal complementarity weight $\mathcal{E}$6 is removed, MRR drops to 38.76. When the inter-modal complementarity weight $\mathcal{E}$7 is removed, MRR drops to 37.64. The paper interprets this as indicating that both levels matter and that inter-modal complementarity is especially important (&&&^^^^2query2^^^^&&&).

The fusion-strategy comparison is also explicit. Concatenation yields MRR 22.34, Hit@^^^^2id:(Li, 28 Jul 2025) OR \2^^^^ 9.57, and Hit@^^^^2id:(Li, 28 Jul 2025) OR \2query2^^^^ 5^^^^2query2^^^^.2^^^^2id:(Li, 28 Jul 2025) OR \2^^^^; element-wise product yields 3^^^^2id:(Li, 28 Jul 2025) OR \2^^^^.96, 23.42, and 5^^^^2id:(Li, 28 Jul 2025) OR \2^^^^.44; gate yields 38.^^^^2query2^^^^5, 28.76, and 54.^^^^2query2^^^^9; and MoCME/CMKF yields 39.62, 29.7^^^^2id:(Li, 28 Jul 2025) OR \2^^^^, and 55.36 (&&&^^^^2query2^^^^&&&). The paper further reports that replacing the Complementary Mixture of Experts backbone with an MLP reduces MRR to 37.63, and replacing it with a linear backbone reduces MRR to 36.84. It also reports best performance when the number of experts is close to the number of modalities (&&&^^^^2query2^^^^&&&).

In a broader research context, CMKF belongs to a family of multimodal methods that treat complementarity as a first-class design principle rather than a by-product of fusion. Earlier work on multimodal deep learning proposed multiplicative combination and multiplicative selection over modality mixtures to handle sample-dependent modality reliability (Liu et al., 2018). Later work in cross-modality object detection fused RGB and infrared features in a hidden state space through a Fusion-Mamba block with State Space Channel Swapping and Dual State Space Fusion (Dong et al., 2024). In modality-agnostic semantic segmentation, CHARM explicitly opposed homogenization and instead proposed collaborative harmonization through a Mutual Perception Unit and dual-path optimization (Wen et al., 5 Aug 2025). Related principles also appear in complementarity-aware multimodal fusion for diffusion-policy imitation learning, where modality-wise dropout is the primary regularizer and cross-attention is a lightweight interaction layer (&&&^^^^2id:(Li, 28 Jul 2025) OR \22^^^^&&&), and in camera-radar fusion via multi-level knowledge distillation that preserves intrinsic modality characteristics while amplifying complementary strengths (&&&^^^^2id:(Li, 28 Jul 2025) OR \2^^^^3&&&).

This suggests that CMKF is part of a wider shift from importance-weighted fusion toward complementarity-aware fusion. Its specific form, however, is narrower and more technical than a generic multimodal recipe. It is a two-stage, information-theoretic fusion module for MMKGC, with structural KG information treated as a modality, expert-based multi-view decomposition inside each modality, mutual-information-derived weights at both fusion stages, and a RotatE-based scoring interface (&&&^^^^2query2^^^^&&&). It is therefore best understood not as a universal synonym for multimodal fusion, but as a particular architecture for exploiting redundancy-free signals in multimodal knowledge graphs.