Pith. sign in

REVIEW 3 major objections 5 minor 46 references

ADMC: Attention-based Diffusion Model for Missing Modalities Feature Completion

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A diffusion network with attention can fill missing modalities in multimodal recognition, the paper claims, and it reports consistent gains on two benchmarks.

desk verdict A useful missing-modality pipeline with strong benchmark numbers, but the diffusion initialization contradicts the paper's own forward process and the central generation claim is not supported. read the letter →

arxiv 2507.05624 v2 pith:ENWNQPY3 submitted 2025-07-08 cs.AI

Yuhan Li , Wei Zhang , Juan Chen , Jiangjia Yan , Peng Xiangli , Liangze Yin This is my paper
classification cs.AI
keywords missingmodalitycompletionmultimodalemotionrecognitionintentdiffusionmodelself-attentionfusionfeaturegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes ADMC, a three-stage framework for completing missing modalities in multimodal emotion and intent recognition. It claims that separately trained per-modality feature extractors avoid the over-coupling that hurts prior joint-training methods, and that an attention-based diffusion network can generate missing modality features that match the distribution of real multimodal data. The paper reports gains over existing completion methods on the IEMOCAP and MIntRec benchmarks, including an average improvement of 1.9% on IEMOCAP and 9.4% on MIntRec, and shows that the same generation step can also improve full-modality recognition. If correct, the framework would give a single model that handles any missing-modality pattern while still performing well when all modalities are present.

What carries the argument

The load-bearing object is the Attention-based Diffusion Network (ADN): a denoising diffusion model whose noise-prediction network is a multi-head self-attention encoder trained on the stacked audio-visual-text features $F^0_{tva}$. Forward diffusion follows $q(F^n_{tva}|F^0_{tva}) = \mathcal{N}(F^n_{tva}; \sqrt{\bar{\alpha}_n}F^0_{tva}, (1-\bar{\alpha}_n)\epsilon)$, and reverse generation uses the standard denoising update starting at step $n$. The masked update in Eq. (4) keeps the available modalities' forward-noised values exactly while regenerating the missing part, so the self-attention over the modality dimension is what lets one modality condition the generation of another. The experimentally chosen starting step $n$ controls how much of the available signal survives into the completed representation.

What would settle it

On a held-out third dataset, vary the starting step $n$ systematically and measure the distributional distance between completed and ground-truth features; if no $n$ yields completed features closer to ground truth than zero-filling, or if the optimal $n$ must be re-tuned for every missing-modality pattern for the reported gains to persist, the conditional-completion assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that poor missing-modality performance comes from coupling features to the classifier and from imprecise generation, and that both can be fixed by decoupling. ADMC first trains each modality encoder on single-modality classification and freezes it, then trains an Attention-based Diffusion Network (ADN) on complete multimodal feature stacks to predict diffusion noise with a multi-head self-attention encoder. At test time, missing entries are replaced by Gaussian noise and the reverse diffusion process, starting from an experimentally chosen step $n$ and masked at each denoising step with the update $F^{n-1} = m\odot F_o^{n-1} + (1-m)\odot F_{tva}^{n-1}$, produces the missing features while leaving available features intact. The paper claims this yields the best results across all six single- and paired-modality missing scenarios on both datasets, and that in the full-modality case the MMER variant, which concatenates an original modality with its cross-modal reconstruction, beats both joint training and frozen encoders without enhancement.

Load-bearing premise

The framework assumes that a denoiser trained on unconditionally noised full-modality data, when started at a hand-tuned noise step $n$ and masked along the way, produces missing features that follow the true conditional distribution given the available modalities.

Editorial extensions

If this is right

  • A single trained ADN, rather than one reconstruction model per missing pattern, can serve all six missing-modality configurations on a dataset.
  • Decoupling modality encoders from the classifier removes the sharp performance drop that prior jointly trained methods show when a modality is absent.
  • Cross-modal generation adds usable signal in the full-modality case: concatenating a modality with its generated counterpart improves recognition over using only original features.
  • The diffusion module adds roughly 50 ms per sample at 1,000 denoising steps on an RTX 4090, and standard diffusion acceleration techniques can cut the steps to 50-100 with little performance loss.
  • In the random missing-rate experiments, the method degrades least as the missing ratio grows from 0.1 to 0.5.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The starting noise step $n$ is a free parameter that is tuned experimentally after training; if the optimal $n$ must be re-chosen per dataset or per missing-modality pattern, the 'one model for all scenarios' advantage weakens.
  • The claim that ADN generates features matching the authentic distribution is tested only indirectly through downstream accuracy and t-SNE plots; a direct distributional measure in feature space, such as MMD or a feature-level FID, would make the mechanism falsifiable beyond benchmark numbers.
  • The MMER result suggests diffusion-based cross-modal reconstruction can act as test-time data augmentation; sampling multiple completions per missing modality and averaging the predictions is a natural extension that the paper does not explore.
  • The framework should transfer to other tri-modal tasks or even paired-modality settings, with the caveat that datasets with one dominant modality or strongly correlated modalities would stress-test whether self-attention can learn useful cross-modal conditioning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes ADMC, a three-stage framework for multimodal emotion and intent recognition under missing modalities. Stage 1 trains separate feature extraction networks for audio, visual, and text modalities via single-modality classification. Stage 2 trains an Attention-based Diffusion Network (ADN) on full-modality latent features to learn a denoiser that captures inter-modal dependencies. At inference, missing modality features are replaced with Gaussian noise and the reverse diffusion process is run from an experimentally chosen starting step n, with the available modalities re-injected through a mask. Stage 3 trains a self-attention-based multimodal fusion network for classification. The paper reports state-of-the-art results on IEMOCAP and MIntRec across all missing-modality scenarios and in the full-modality case via a cross-modal enhancement method (MMER), with an average performance improvement of 9.4% on MIntRec.

Significance. If the results are valid, the proposed architecture offers a practical and modular recipe for missing-modality robustness, combining independently trained unimodal encoders with a diffusion-based generator. The design choice of decoupling feature extraction from completion is well motivated, and the ablation study provides useful evidence for the contribution of each component. The paper also ships a concrete latency analysis for the diffusion module. However, the central generation mechanism has a technical inconsistency in the reverse-process initialization, and the evaluation protocol lacks statistical rigor (no error bars, mixed reproduced and originally reported baselines). These issues prevent the empirical claims from being verified as stated.

major comments (3)
  1. [§3.2–3.3 (reverse process initialization)] The reverse process is not initialized consistently with the forward process. With the reported linear schedule beta_s=0.0003, beta_e=0.06, N=1000, at n=50 the cumulative alpha_bar_50 is approximately 0.91, so the forward marginal q(F^50|F^0) is a slightly noised version of the data with per-dimension variance around 0.09, not N(0,I). The missing channels are initialized from N(0,I) (Section 3.3), placing them far outside the denoiser's training distribution. Equation (4) then re-injects available channels at their correct forward-noise level, so the self-attention network processes a tensor whose channels have inconsistent noise levels throughout the reverse trajectory. Consequently, the claim that ADN generates missing features 'closely aligned with the authentic multimodal distribution' is not supported by the presented sampling procedure. The authors should either (a) start the reverse process from a step where q(F^n) is close to N(0,I), (b) provide a corrected initialization that samples the missing channel from the appropriate conditional (or marginal) distribution at step n, or (c) empirically verify that the generated features are drawn from the learned conditional distribution (e.g., by comparing distributions with FID/MMD and showing that the downstream fusion accuracy matches that of ground-truth features). Without this, the reported gains could be an artifact of the fusion network exploiting available-modality statistics.
  2. [§5.1 (comparative study) and Tables 1–4] The evaluation does not report error bars or multiple seeds, and the baseline results are a mix of numbers reproduced from open-source code (marked *) and numbers taken from the original papers. Because the train/validation splits, feature extractors, and random seeds may differ, the observed improvements (e.g., 'an average performance improvement of 9.4%' on MIntRec) cannot be distinguished from experimental variation. At least for the key comparisons, the authors should report mean and standard deviation over several runs, and ideally re-run all baselines under the same protocol with the same feature sets. This is essential to support the state-of-the-art claim.
  3. [§5.3 (visualization analysis)] The choice of the reverse starting step n is load-bearing, yet it is justified only by a qualitative t-SNE plot (Figure 3). The paper states that 'the optimal n is determined experimentally' but provides no quantitative sensitivity analysis, no per-dataset or per-scenario values, and no measure of distributional similarity between generated and ground-truth features. Because the consistency of the reverse process depends on n (see the initialization issue above), the manuscript should report accuracy or a distributional metric as a function of n, and disclose the selected n for each setting. In addition, the actual value of alpha_bar_n should be reported so that the reader can assess the noise-level mismatch.
minor comments (5)
  1. [Eq. (1)] The notation epsilon is used as a variance parameter: q(F^n_tva | F^0_tva) = N(..., (1 - alpha_bar_n) epsilon). This should be an identity covariance matrix; as written, the variance is a random variable, and epsilon is already used for the noise. Please fix.
  2. [Table 1] In the 'Ours' row, the {AV} and {AT} entries are concatenated as '72.1/72.974.1/76.6'; insert a space or separator.
  3. [§5.5] The sentence 'we using a probabilistic approach' is ungrammatical; also 'p∈0.1,0.2,...,0.5' should be presented as a set, e.g., p ∈ {0.1, 0.2, ..., 0.5}.
  4. [References] References [3] and [4] are the same reference (Cai et al. 2018), duplicated with different numbers; renumber accordingly and remove the duplicate.
  5. [§5.2 and Tables 3–4] The label 'w/ U-Net' in the ablation tables is confusing; the text says 'Replacing the self-attention module in ADN with a standard U-Net,' so the label should be something like 'U-Net instead of attention' to avoid implying the full model includes U-Net in addition to the attention module.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found: the ADMC pipeline is trained on full-modality data and evaluated against external benchmarks, with standard validation-set model selection.

full rationale

Walking the paper's derivation chain, the load-bearing components are not defined in terms of their own outputs. The ADN is trained on full-modality features with a standard DDPM-style objective (Eq. 1 and Eq. 2), and the reverse generation in Eq. 3 with the available-modality replacement in Eq. 4 is a conditional completion mechanism; the downstream fusion network in Eq. 6 is trained separately. None of these equations is constructed from the target benchmark metrics or from the claimed SOTA improvements. The reported results are empirical comparisons against external baselines (CRA, MCTN, MMIN, IF-MMIN, IMDer, MPLMM, P-RMF) that come from other groups, and the ablation study isolates the contributions of ITFN and ADN. The starting diffusion step n is selected experimentally and visualized with t-SNE (Section 5.3); this is hyperparameter selection on the validation set rather than a fitted parameter being renamed as a prediction. The t-SNE evidence is qualitative and the reverse-process initialization at n=50 may be technically questionable, but those are correctness or evidence-quality concerns, not circularity in the sense of a prediction reducing to its inputs by construction. There is also no load-bearing self-citation chain: the cited feature-extraction conventions [36,37,41] are external prior work, not the authors' own uniqueness or optimality claims. Therefore the central derivation is self-contained against external benchmarks, and the paper merits a low circularity score.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The main free parameter is the reverse diffusion starting step n, which is chosen experimentally and directly controls the generated features. The core axioms are standard DDPM theory and the domain assumption that an unconditionally trained attention network, combined with mask-based replacement, yields valid conditional samples for the missing modalities. No new entities are introduced.

free parameters (1)
  • Reverse diffusion starting step n = Not reported; visualization in Fig. 3 favors n=50
    Section 3.2: 'The optimal n is determined experimentally.' Section 5.3: 'When n = 50, the generated embeddings closely align...' Performance in Tables 1-4 depends on this choice; if tuned per scenario, it is a free parameter fit to the validation set.
assumptions (4)
  • standard math DDPM forward and reverse processes as defined in Eqs. 1-3 (Gaussian noising and learned denoising) are valid for the concatenated three-modality feature vectors.
    The method relies on standard diffusion model theory from Ho et al. [11] and Song et al. [26]. No modification to the mathematical framework is attempted.
  • domain assumption An unconditionally trained denoising network, combined with mask-based replacement of available modalities (Eq. 4), produces missing-modality samples from the correct conditional distribution p(missing | available).
    Section 3.2-3.3: reverse generation starts from an experimentally chosen step n and injects available modalities via a mask. The paper provides no formal argument or evaluation that these samples follow the true conditional distribution; the only evidence is the t-SNE visualization (Fig. 3) and downstream accuracy.
  • domain assumption Features extracted by the independently trained single-modality networks are informative enough to support later diffusion and fusion when modalities are missing.
    Section 3.1: encoders are trained on single-modality classification with early stopping and then fixed. The w/o ITFN ablation supports this, but the assumption that decoupled features do not lose cross-modal information needed by the diffusion model is not proven.
  • ad hoc to paper A single experimentally chosen reverse starting step n generalizes across all six missing-modality scenarios and both datasets.
    Section 3.2 says n is determined experimentally; Section 5.3 suggests n=50 is used. If n is instead tuned per scenario, the generalization assumption is violated and the reported numbers are partially fitted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADMC: Attention-based Diffusion Model for Missing Modalities Feature Completion." pith.science (2026). https://pith.science/paper/ENWNQPY3

@misc{pith2026250705624,
  author       = {Pith},
  title        = {Pith review of: ADMC: Attention-based Diffusion Model for Missing Modalities Feature Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ENWNQPY3}},
  note         = {Machine review of arXiv:2507.05624}
}
read the original abstract

Multimodal emotion and intent recognition is essential for automated human-computer interaction, It aims to analyze users' speech, text, and visual information to predict their emotions or intent. One of the significant challenges is that missing modalities due to sensor malfunctions or incomplete data. Traditional methods that attempt to reconstruct missing information often suffer from over-coupling and imprecise generation processes, leading to suboptimal outcomes. To address these issues, we introduce an Attention-based Diffusion model for Missing Modalities feature Completion (ADMC). Our framework independently trains feature extraction networks for each modality, preserving their unique characteristics and avoiding over-coupling. The Attention-based Diffusion Network (ADN) generates missing modality features that closely align with authentic multimodal distribution, enhancing performance across all missing-modality scenarios. Moreover, ADN's cross-modal generation offers improved recognition even in full-modality contexts. Our approach achieves state-of-the-art results on the IEMOCAP and MIntRec benchmarks, demonstrating its effectiveness in both missing and complete modality scenarios.

Figures

Figures reproduced from arXiv: 2507.05624 by the authors.

Figure 1
Figure 1. In real-world human-computer interaction, modal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Shows the overall framework of the proposed ADMC method. (a) The independently trained feature extraction [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Results under different missing rates on the IEMO [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Shows the visualization of ground-truth multimodal [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 26 canonical work pages

  1. [1]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems33 (2020), 12449–12460

    2020

  2. [2]

    Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. 2008. IEMOCAP: Interactive emotional dyadic motion capture database.Language resources and evaluation42 (2008), 335–359

    work page 2008

  3. [4]

    Lei Cai, Zhengyang Wang, Hongyang Gao, Dinggang Shen, and Shuiwang Ji

  4. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv. arXiv preprint arXiv:1810.04805(2019)

    arXiv 2019

  5. [6]

    InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining

    Deep adversarial learning for multi-modality missing data completion. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1158–1166

    work page

  6. [7]

    Y. Duan, Y. Lv, W. Kang, and Y. Zhao. 2014. A deep learning based approach for traffic data imputation. InProceedings of the 17th International IEEE Conference on Intelligent Transportation Systems (ITSC). IEEE, 912–917

    work page 2014

  7. [8]

    Changde Du, Changying Du, Hao Wang, Jinpeng Li, Wei-Long Zheng, Bao- Liang Lu, and Huiguang He. 2018. Semi-supervised deep generative modelling Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yuhan Li, Wei Zhang, Juan Chen, Jiangjia Yan, Peng Xiangli, and Liangze Yin of incomplete multi-modality emotional data. InProceedings of the 26th ACM intern...

    work page 2018

  8. [9]

    Haiyan Fan, Yunjin Chen, Yulan Guo, Hongyan Zhang, and Gangyao Kuang

    work page

Show all 46 references
  1. [10]

    Florian Eyben, Martin Wöllmer, and Björn Schuller. 2010. Opensmile: the munich versatile and fast open-source audio feature extractor. InProceedings of the 18th ACM international conference on Multimedia. 1459–1462

    2010

  2. [11]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  3. [12]

    Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598(2022)

    2022 arXiv

  4. [13]

    Z. Guo, T. Jin, and Z. Zhao. 2024. Multimodal prompt learning with miss- ing modalities for sentiment analysis and emotion recognition.arXiv preprint arXiv:2407.05374(2024)

    2024 arXiv

  5. [14]

    Xuejian Huang, Tinghuai Ma, Li Jia, Yuanjian Zhang, Huan Rong, and Najla Alnabhan. 2023. An effective multimodal representation and fusion method for multimodal intent recognition.Neurocomputing548 (2023), 126373

    2023

  6. [15]

    Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 1746–1751

    2014

  7. [16]

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger

  8. [17]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition. 4700–4708

  9. [18]

    PP Liang, Z Liu, YHH Tsai, Q Zhao, R Salakhutdinov, and LP Morency. 2019. Learning representations from imperfect time series data via tensor rank regu- larization. arXiv 2019.arXiv preprint arXiv:1907.01011(2019)

    2019 arXiv

  10. [19]

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. 2023. AudioLDM: Text-to-Audio Generation with Latent Diffusion Models.arXiv preprint arXiv:2301.12503(2023)

    2023 arXiv

  11. [20]

    Zaijing Li, Fengxiao Tang, Ming Zhao, and Yusen Zhu. 2022. EmoCaps: Emotion capsule based model for conversational emotion recognition.arXiv preprint arXiv:2203.13504(2022)

    2022 arXiv

  12. [21]

    Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. 2023. GCNet: Graph completion network for incomplete multimodal learning in conversation.IEEE Transactions on pattern analysis and machine intelligence45, 7 (2023), 8419–8432

    2023

  13. [22]

    Hai Pham, Paul Pu Liang, Thomas Manzini, Louis-Philippe Morency, and Barn- abás Póczos. 2019. Found in translation: Learning robust joint representations by cyclic translations between modalities. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 6892–6899

    2019

  14. [23]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention. Springer, 234–241

    2015

  15. [24]

    Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffu- sion probabilistic models. InInternational conference on machine learning. PMLR, 8162–8171

    2021

  16. [25]

    Srinivas Parthasarathy and Shiva Sundaram. 2020. Training strategies to handle missing modalities for audio-visual expression recognition. InCompanion Publi- cation of the 2020 International Conference on Multimodal Interaction. 400–404

    2020

  17. [26]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. arXiv:2011.13456 [cs.LG] https://arxiv.org/abs/2011. 13456

    2021 arXiv

  18. [27]

    Luan Tran, Xiaoming Liu, Jiayu Zhou, and Rong Jin. 2017. Missing Modalities Imputation via Cascaded Residual Autoencoder. InComputer Vision & Pattern Recognition. 4971–4980

    2017

  19. [28]

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 conference proceedings. 1–10

    2022

  20. [29]

    Haşim Sak, Andrew Senior, and Françoise Beaufays. 2014. Long short-term mem- ory based recurrent neural network architectures for large vocabulary speech recognition.arXiv preprint arXiv:1402.1128(2014)

    2014 arXiv

  21. [30]

    Yuanzhi Wang, Yong Li, and Zhen Cui. 2023. Incomplete Multimodality-Diffused Emotion Recognition.Advances in Neural Information Processing Systems36 (2023), 17117–17128

    2023

  22. [31]

    Zilong Wang, Zhaohong Wan, and Xiaojun Wan. 2020. Transmodality: An end2end fusion method with transformer for multimodal sentiment analysis. In Proceedings of the web conference 2020. 2514–2520

    2020

  23. [32]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing Data using t-SNE.Journal of Machine Learning Research9 (2008), 2579–2605

    2008

  24. [33]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  25. [34]

    Jiandian Zeng, Jiantao Zhou, and Tianyi Liu. 2022. Mitigating inconsistencies in multimodal sentiment analysis under uncertain missing modalities. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2924– 2934

    2022

  26. [35]

    Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. 2020. Deep partial multi-view learning.IEEE transactions on pattern analysis and machine intelligence44, 5 (2020), 2402–2415

    2020

  27. [36]

    Ziqi Yuan, Wei Li, Hua Xu, and Wenmeng Yu. 2021. Transformer-based feature reconstruction network for robust multimodal sentiment analysis. InProceedings of the 29th ACM International Conference on Multimedia. 4400–4407

    2021

  28. [37]

    Jiandian Zeng, Tianyi Liu, and Jiantao Zhou. 2022. Tag-assisted multimodal sentiment analysis under uncertain missing modalities. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1545–1554

    2022

  29. [38]

    Shanshan Zhong, Zhongzhan Huang, Weushao Wen, Jinghui Qin, and Liang Lin

  30. [39]

    Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. 2024. Token-level contrastive learning with modality-aware prompting for multimodal intent recognition. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 17114–17122

    2024

  31. [40]

    Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng

  32. [41]

    Haolin Zuo, Rui Liu, Jinming Zhao, Guanglai Gao, and Haizhou Li. 2023. Ex- ploiting modality-invariant feature for robust multimodal emotion recognition with missing modalities. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP...

    2023

  33. [42]

    Jinming Zhao, Ruichen Li, and Qin Jin. 2021. Missing modality imagination net- work for emotion recognition with uncertain missing modalities. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference ...

    2021

  34. [46]

    Aoqiang Zhu, Min Hu, Xiaohua Wang, Jiaoyun Yang, Yiming Tang, and Ning An

  35. [2017]

    Hyperspectral image restoration using low-rank tensor recovery.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing10, 10 (2017), 4589–4604

    2017

  36. [2022]

    InProceedings of the 30th ACM International Conference on Multimedia

    Mintrec: A new dataset for multimodal intent recognition. InProceedings of the 30th ACM International Conference on Multimedia. 1688–1697

  37. [2023]

    InProceedings of the 31st ACM International Conference on Multimedia

    Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. InProceedings of the 31st ACM International Conference on Multimedia. 567–578

  38. [2025]

    InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Proxy-Driven Robust Multimodal Sentiment Analysis with Incomplete Data. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 22123–22138

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.