Pith. sign in

REVIEW 2 major objections 5 minor 45 references

A self-supervised model learns transferrable EEG emotion features by combining masked prediction with stimulus-aligned contrastive learning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 21:24 UTC pith:GUBZLR7I

load-bearing objection Solid empirical SSL recipe for cross-dataset affective EEG; gains real but modest, driven by a same-stimulus contrastive bias that is plausible yet untested against residual subject confounds. the 2 major comments →

arxiv 2607.04139 v1 pith:GUBZLR7I submitted 2026-07-05 cs.LG

Masked Generative-Contrastive Representation Learning for Cross-Dataset EEG-Based Emotion Recognition

Huqin Weng , Jiayang Huang , Yimin Wen , Jie Du , Chi-Man Vong , Chuangquan Chen This is my paper
classification cs.LG
keywords EEG emotion recognitionself-supervised learningcross-dataset transfermasked generative learningcontrastive learningregion-aware graph convolutionJEPAcross-subject generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

EEG emotion recognition models usually fail when the recording device, electrode layout, or subject changes. This paper claims that pretraining a single encoder on a large unlabeled set (FACED) and then fine-tuning it on smaller labeled sets (SEED-IV/V/VII) can produce subject- and device-invariant emotion features if three design choices are made together. First, electrodes are grouped into fixed brain regions and processed by graph convolution so that different channel counts still yield comparable regional features. Second, a JEPA-style generative head predicts latent embeddings of masked temporal blocks rather than raw signal values, yielding noise-robust fine-grained features. Third, a contrastive head treats both the masked and unmasked views of the same segment as strong positives and, more importantly, treats segments elicited by the same video stimulus (even from different subjects) as dynamically weighted positives. The resulting representations transfer better than strong self-supervised baselines under leave-one-subject-out evaluation, both within and across recording sessions.

Core claim

MGCRL learns universal EEG emotion representations by jointly optimizing a region-aware spatiotemporal encoder, a JEPA latent-prediction generative loss, and a masked dynamic contrastive loss that exploits temporal stability and cross-subject similarity under identical stimuli; when pretrained on FACED and fine-tuned on SEED-series data under identical hyperparameters, it outperforms competitive SSL baselines on cross-subject emotion recognition.

What carries the argument

The masked dynamic contrastive module: masked and original features of the same sample form strong positives, while a stimulus-label mask together with a feature-derived dynamic weight matrix treats same-stimulus segments (across time or subjects) as soft positives, re-estimated every iteration.

Load-bearing premise

Segments recorded while different people watch the same video clip share enough stable, emotion-relevant content that treating them as positive pairs improves, rather than confuses, the learned representation.

What would settle it

Replace the stimulus-based positive pairs with random or subject-matched pairs of equal size; if the accuracy gains on SEED-IV/V/VII vanish or reverse, the contrastive premise is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Pretraining once on a large multi-subject EEG corpus can reduce the amount of labeled data needed for a new emotion recognition device or subject pool.
  • Region-based graph convolution can absorb channel-layout mismatches without discarding non-overlapping electrodes.
  • Latent-space prediction (JEPA) is preferable to raw-signal reconstruction for noisy EEG.
  • Stimulus-aligned contrastive sampling can be reused for any affective or cognitive task that shares elicitation material across subjects.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same-stimulus assumption holds only for highly arousing clips, the method may degrade on milder or continuous-label emotion corpora.
  • The same dual-use masking (context for JEPA + augmentation for contrastive) could be applied to other biosignals that share event markers across subjects, such as fNIRS or EMG.
  • Because the contrastive term is retained at fine-tuning, the framework may also improve purely supervised EEG pipelines that already have limited labels.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The manuscript proposes MGCRL, a self-supervised pretraining framework for cross-dataset EEG emotion recognition. It combines a region-aware spatiotemporal encoder (temporal convolution + Intra-Brain-Region GCN with hybrid physical–dynamic adjacency on a 9×9 padded grid, followed by temporal multi-head attention), a JEPA-style generative branch that predicts latent embeddings of masked temporal blocks, and a masked dynamic contrastive branch that treats masked/unmasked views of the same segment as strong positives and same-stimulus segments (across time and subjects) as soft positives weighted by current cosine similarity (Eqs. 16–19). Pretraining is performed on unlabeled FACED; the encoder is fine-tuned with retained contrastive loss plus cross-entropy on SEED-IV/V/VII under within- and cross-session leave-one-subject-out protocols. Tables I and IV report consistent gains over BIOT, TS-TCC, EEG2Rep and LaBraM (relative improvements of roughly 3–4 points), with ablations (Tables II–III) attributing the largest drop to removal of the contrastive term.

Significance. Cross-dataset, cross-subject EEG emotion recognition under heterogeneous channel counts and label sets remains a genuine bottleneck for affective BCIs. The dual-use masking design that unifies JEPA-style latent prediction with stimulus-aware contrastive learning is a clean architectural contribution, and the decision to keep identical hyperparameters across three target datasets and both within- and cross-session LOSO is a methodological strength that few EEG SSL papers match. If the same-stimulus inductive bias is shown to be primarily emotion-driven rather than subject-confounded, the framework would supply a practical, label-efficient pretraining recipe transferable beyond the SEED family. Absolute accuracies remain modest and standard deviations large, so the practical impact is incremental rather than transformative, but the empirical protocol is solid enough to be useful to the community.

major comments (2)
  1. §III.C.3, Eqs. (16)–(19) and Table II: the largest performance drop (3.6–6.9 points) occurs when L_contrast is removed, so the claimed superiority over EEG2Rep/LaBraM rests almost entirely on the same-stimulus soft-positive construction. The paper never quantifies residual subject or session confounds inside those pairs (e.g., by reporting within-stimulus subject-ID classification accuracy on the learned embeddings, or by an ablation that replaces stimulus labels with random or subject-matched negatives). Without such a control, it remains possible that the contrastive term is largely aligning subject idiosyncrasies that co-occur with the same video rather than emotion-invariant content; the cross-subject LOSO numbers would then be optimistic. A short diagnostic experiment or explicit discussion of this risk is needed to underwrite the central claim of “universal representation.”
  2. Tables I and IV: absolute accuracies are low (≈35–41 % on 4–7-class tasks) and standard deviations are large (often 6–7 %). While relative gains over baselines are consistent, the manuscript should report statistical significance (paired tests across subjects) and chance-level baselines so that readers can judge whether the improvements are reliable rather than noise. The claim of “stable and superior results” under identical hyperparameters is otherwise only weakly supported by the reported means.
minor comments (5)
  1. Abstract and §I: “noise robustness fine-grained representations” is ungrammatical; revise to “noise-robust fine-grained representations.”
  2. Fig. 2 caption and §III.B: the three-stage pipeline is clear, but the precise flow of gradients between the EMA target encoder and the contrastive projector is not stated; a short sentence would help reproducibility.
  3. §IV.B: free parameters (K, mask-ratio ranges, η, τ, embedding dim, Transformer depth) are listed but no sensitivity analysis is provided; even a brief note that results are stable within the chosen ranges would strengthen confidence.
  4. Fig. 6: brain-region weight visualizations are informative, yet the ranking procedure (average of K kernels) is only briefly described; stating whether weights are L1-normalized or absolute would improve interpretability.
  5. Throughout: minor typographical inconsistencies (“tr ansfer”, “glob al”, “su bject-invariant”) remain from line-breaking; a final proof-read is needed.

Circularity Check

0 steps flagged

Empirical SSL pipeline with no by-construction reduction of reported accuracies; stimulus-positive contrastive bias is inductive, not tautological.

full rationale

MGCRL is a standard pretrain-then-fine-tune SSL method. Pretraining optimizes the sum of an L2 JEPA latent-prediction loss (Eq. 12) and a masked dynamic contrastive loss (Eqs. 15–19) on unlabeled FACED segments; fine-tuning adds cross-entropy on labeled SEED data and evaluates leave-one-subject-out accuracy on held-out subjects/sessions (Tables I, IV). No equation equates a reported accuracy number to a fitted constant or to an input definition. The same-stimulus mask M_stim (Eq. 16) and dynamic weights D_weights (Eq. 17) encode an inductive bias that segments sharing a video label are soft positives; that bias can be false or confounded, but it is not circular—the loss still has to be optimized and the resulting encoder still has to transfer. Ablations (Table II) and baseline comparisons are ordinary empirical controls, not self-referential derivations. Self-citations are limited to architectural building blocks (JEPA, SimCLR, region partitions) that are independently published; none is a uniqueness theorem that forces the central claim. Consequently the circularity score is 1 (minor self-citation of standard components, no load-bearing circular step).

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central transfer claim rests on a handful of modeling choices and free hyperparameters rather than on theorems. The free parameters control capacity and optimization; the domain axioms encode neuroscience priors (region partitions, distance-based adjacency, same-stimulus emotional similarity) that are taken from prior literature or asserted without new validation; the invented entities are the architectural modules themselves, which have no independent existence outside this paper’s experiments.

free parameters (5)
  • temporal-convolution kernel count K = 40
    Set to 40 following EEG Conformer; directly determines feature-map width and capacity of the spatiotemporal encoder.
  • context / predictor mask-ratio ranges = (0.85–1.0) / (0.15–0.2)
    Randomly sampled from (0.85,1.0) and (0.15,0.2) respectively; control the difficulty of the JEPA task and the strength of the contrastive views.
  • pretrain / finetune learning rates and epochs = 0.01 / 0.0001, 100 ep
    0.01 / 0.0001, 100 epochs each, batch 64; chosen by the authors and kept fixed across all SEED targets.
  • EMA decay η and contrastive temperature τ
    Control target-encoder update speed and softness of the contrastive distribution; values not exhaustively ablated.
  • embedding dimension and Transformer depth / heads = 128 / 3 / 8
    128-dim, depth 3, 8 heads; set “in line with JEPA” and kept constant.
axioms (5)
  • domain assumption Partitioning the scalp into the 17 fixed brain regions of Song et al. (2020) yields functionally meaningful subgraphs for emotion-related EEG.
    Invoked in Section III.C.1 and Fig. 3; the IBR-GCN operates only inside these regions.
  • domain assumption Physical electrode adjacency can be approximated by the inverse-square of spatial distance, and a trainable residual Δ can refine it.
    Eq. (1); taken from prior GCN-EEG literature without new physiological validation.
  • ad hoc to paper EEG responses to the identical video stimulus share temporally stable and cross-subject-invariant emotional content that can be treated as positive pairs.
    Core of the masked dynamic contrastive module (Eqs. 16–19); the paper’s largest ablation gain depends on it.
  • domain assumption Latent-space prediction (JEPA) is preferable to signal-level reconstruction for noisy EEG because it avoids fitting irrelevant low-level detail.
    Stated in Related Work and Section III.C.2; inherited from I-JEPA without EEG-specific theoretical guarantee.
  • ad hoc to paper Zero-padding missing electrodes on a 9×9 grid plus region-wise masking preserves enough spatial information for cross-dataset transfer.
    Formulation paragraph and Fig. 3; the entire channel-mismatch solution rests on this construction.
invented entities (3)
  • Intra-Brain-Region Graph Convolution (IBR-GCN) with hybrid physical–dynamic adjacency no independent evidence
    purpose: Capture localized spatial and sample-specific functional relations while tolerating missing channels.
    Defined by Eqs. (1)–(3); no independent physiological or multi-lab validation outside this paper’s ablations.
  • Masked Dynamic Contrastive Learning (MDCL) with stimulus-driven soft positives no independent evidence
    purpose: Simultaneously enforce intra-sample view consistency and inter-subject stimulus invariance.
    Eqs. (15)–(19); the dynamic weight matrix is recomputed each iteration from the model’s own embeddings.
  • MGCRL dual-use masking pipeline no independent evidence
    purpose: Let the same random mask serve both as JEPA context and as contrastive augmentation.
    Architectural unification claimed in the introduction; evaluated only inside the authors’ experimental suite.

pith-pipeline@v1.1.0-grok45 · 23066 in / 3714 out tokens · 37481 ms · 2026-07-11T21:24:31.986584+00:00 · methodology

0 comments
read the original abstract

Self-supervised learning (SSL) shows strong potential for cross-dataset transfer by improving feature representation and generalization. However, its application to EEG-based emotion recognition remains largely unexplored. Existing SSL methods struggle to capture the intricate spatiotemporal dependencies of EEG signals under varying channel configurations, extract fine-grained representations resilient to noise, and derive global features that generalize well across subjects. To address these challenges, we propose Masked Generative-Contrastive Representation Learning (MGCRL), a novel SSL framework specifically designed for EEG-based emotion recognition. Built upon a region-aware spatiotemporal encoder, MGCRL integrates generative and contrastive learning to achieve both fine-grained and global discriminative representations for cross-dataset generalization. MGCRL introduces three key designs: 1) a spatiotemporal encoder that incorporates region-based graph convolution to capture localized spatial and functional relationships, enhancing region-specific feature learning and mitigating the impact of varying EEG channel configurations across datasets; 2) a generative learning mechanism based on the joint embedding predictive architecture (JEPA) that utilizes masked features to capture noise robustness fine-grained representations, improving the model's capability to characterize subtle emotional states; and 3) a contrastive learning strategy that leverages masked and original features to learn temporally stable and cross-subject-invariant representations across the same stimuli, boosting emotion discrimination and cross-subject generalization. Under these designs, MGCRL exhibits remarkable ability to learn universal representation. Extensive experiments involving pretraining on the large FACED dataset and fine-tuning on multiple SEED-series datasets demonstrate the effectiveness of MGCRL.

Figures

Figures reproduced from arXiv: 2607.04139 by Chi-Man Vong, Chuangquan Chen, Huqin Weng, Jiayang Huang, Jie Du, Yimin Wen.

Figure 1
Figure 1. Figure 1: Illustration of heterogeneities in cross-dataset problem. These include device heterogeneity, label heterogeneity, and channels heterogeneity [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall structure of MGCRL. (a) MGCRL consists of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: EEG Channel Mapping (9×9 Grid vs. Brain Region Division), where 0 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Experimental results of MGCRL and EEG2Rep on three datasets fo [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The results of fine-tuning different number of subjects on SEED-IV, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The brain region weights are visualized across the thr [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 3 canonical work pages

  1. [1]

    Affective brain–computer interfaces (abcis): A tutorial,

    D. Wu, B.-L. Lu, B. Hu, and Z. Zeng, “Affective brain–computer interfaces (abcis): A tutorial,” Proceedings of the IEEE , vol. 111, no. 10, pp. 1314– 1332, 2023

    2023

  2. [2]

    EEG-based affective brain-computer interfaces: recent advancements and future challenges,

    Y. Chen et al., “EEG-based affective brain-computer interfaces: recent advancements and future challenges,” J. Neural Eng., 2025

    2025

  3. [3]

    A dry electrode for EEG recording,

    B. A. Taheri, R. T. Knight, and R. L. Smith, “A dry electrode for EEG recording,” Electroencephalogr. Clin. Neurophysiol., vol. 90, no. 5, pp. 376–383, 1994

    1994

  4. [4]

    Transfer learning for EEG-based brain–computer interfaces: A review of progress made since 2016,

    D. Wu, Y. Xu, and B.-L. Lu, “Transfer learning for EEG-based brain–computer interfaces: A review of progress made since 2016,” IEEE Trans. Cogn. Dev. Syst., vol. 14, no. 1, pp. 4–19, 2020

    2016

  5. [5]

    Transfer learning in brain- computer interfaces,

    V. Jayaram, M. Alamgir, Y. Altun, B. Scholkopf, and M. Grosse-Wentrup, “Transfer learning in brain- computer interfaces,” IEEE Comput. Intell. Mag. , vol. 11, no. 1, pp. 20–31, 2016

    2016

  6. [6]

    Comprehensive Multisource Learning Network for Cross-Subject Multimodal Emotion Recognition,

    C. Chen et al., “Comprehensive Multisource Learning Network for Cross-Subject Multimodal Emotion Recognition,” IEEE Trans. Emerg. Top. Comput. Intell., vol. 9, pp. 365–380, 2024

    2024

  7. [7]

    A parallel neural networks for emotion recognition based on EEG signals,

    R. He, Y. Jie, W. Tong, M. Zhang, G. Zhu, and E. Q. Wu, “A parallel neural networks for emotion recognition based on EEG signals,” Neurocomputing, vol. 610, p. 128624, 2024, doi: https://doi.org/10.1016/j.neucom.2024.128624

    work page doi:10.1016/j.neucom.2024.128624 2024

  8. [8]

    FCAnet: A novel feature fusion approach to EEG emotion recognition based on cross-attention networks,

    M. Li, H. Huang, and K. Huang, “FCAnet: A novel feature fusion approach to EEG emotion recognition based on cross-attention networks,” Neurocomputing, vol. 638, p. 130102, 2025, doi: https://doi.org/10.1016/j.neucom.2025.130102

    work page doi:10.1016/j.neucom.2025.130102 2025

  9. [9]

    GNN-based multi-source domain prototype representation for cross-subject EEG emotion recognition,

    Y. Guo, C. Tang, H. Wu, and B. Chen, “GNN-based multi-source domain prototype representation for cross-subject EEG emotion recognition,” Neurocomputing, vol. 609, p. 128445, 2024, doi: https://doi.org/10.1016/j.neucom.2024.128445

    work page doi:10.1016/j.neucom.2024.128445 2024

  10. [10]

    Transformer convolutional neural networks for automated artifact detection in scalp EEG,

    W. Y. Peh, Y. Yao, and J. Dauwels, “Transformer convolutional neural networks for automated artifact detection in scalp EEG,” in 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2022, pp. 3599– 3602

    2022

  11. [11]

    LGGNet: Learning from local-global-graph representations for brain–computer interface,

    Y. Ding, N. Robinson, C. Tong, Q. Zeng, and C. Guan, “LGGNet: Learning from local-global-graph representations for brain–computer interface,” IEEE Trans. Neural Netw. Learn. Syst., 2023

    2023

  12. [12]

    CLDTA: Contrastive learning based on diagonal transformer autoencoder for cross-dataset EEG emotion recognition,

    Y. Liao et al. , “CLDTA: Contrastive learning based on diagonal transformer autoencoder for cross-dataset EEG emotion recognition,” arXiv preprint arXiv:2406.08081, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Spp-eegnet: An input-agnostic self-supervised eeg representation model for inter- dataset transfer learning,

    X. Li and V. Metsis, “Spp-eegnet: An input-agnostic self-supervised eeg representation model for inter- dataset transfer learning,” in International Conference on Computing and Information Technology , 2022, pp. 173–182

    2022

  14. [14]

    Time-series representation learning via temporal and contextual contrasting,

    E. Eldele et al. , “Time-series representation learning via temporal and contextual contrasting,” arXiv preprint arXiv:2106.14112, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    Self-supervised contrastive pre-training for time series via time-frequency consistency,

    X. Zhang, Z. Zhao, T. Tsiligkaridis, and M. Zitnik, “Self-supervised contrastive pre-training for time series via time-frequency consistency,” Adv. Neural Inf. Process. Syst., vol. 35, pp. 3988–4003, 2022

    2022

  16. [16]

    Large brain model for learning generic representations with tremendous EEG data in BCI,

    W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, “Large brain model for learning generic representations with tremendous EEG data in BCI,” arXiv preprint arXiv:2405.18765, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    A multi- view spectral-spatial-temporal masked autoencoder for decoding emotions with self-supervised learning,

    R. Li, Y. Wang, W.-L. Zheng, and B.-L. Lu, “A multi- view spectral-spatial-temporal masked autoencoder for decoding emotions with self-supervised learning,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 6–14

    2022

  18. [18]

    Biot: Biosignal transformer for cross-data learning in the wild,

    C. Yang, M. Westover, and J. Sun, “Biot: Biosignal transformer for cross-data learning in the wild,” Adv. Neural Inf. Process. Syst. , vol. 36, pp. 78240–78260, 2023

    2023

  19. [19]

    Eeg2rep: enhancing self- supervised EEG representation through informative masked inputs,

    N. M. Foumani, G. Mackellar, S. Ghane, S. Irtza, N. Nguyen, and M. Salehi, “Eeg2rep: enhancing self- supervised EEG representation through informative masked inputs,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 5544–5555

    2024

  20. [20]

    Contrastive representation learning for electroencephalogram classification,

    M. N. Mohsenvand, M. R. Izadi, and P. Maes, “Contrastive representation learning for electroencephalogram classification,” in Machine Learning for Health, 2020, pp. 238–253

    2020

  21. [21]

    Exploiting multiple eeg data domains with adversarial learning,

    D. Bethge, P. Hallgarten, O. Özdenizci, R. Mikut, A. Schmidt, and T. Grosse-Puppendahl, “Exploiting multiple eeg data domains with adversarial learning,” in 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2022, pp. 3154–3158

    2022

  22. [22]

    Entropy-guided robust feature domain adaptation for electroencephalogram-based cross-dataset drowsiness recognition,

    L. Yuan, J. Cui, R. Li, Z. Zheng, M. Y. Siyal, and Z. Yi, “Entropy-guided robust feature domain adaptation for electroencephalogram-based cross-dataset drowsiness recognition,” Eng. Appl. Artif. Intell. , vol. 137, p. 109153, 2024, doi: https://doi.org/10.1016/j.engappai.2024.109153

    doi:10.1016/j.engappai.2024.109153 2024

  23. [23]

    Generalizable Movement Intention Recognition with Multiple Heterogeneous EEG Datasets,

    X. Gu, J. Han, G.-Z. Yang, and B. Lo, “Generalizable Movement Intention Recognition with Multiple Heterogeneous EEG Datasets,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 9858–9864

    2023

  24. [24]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick, “Masked autoencoders are scalable vision learners,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16000–16009

    2022

  25. [25]

    Self-supervised learning from images with a joint-embedding predictive architecture,

    M. Assran et al. , “Self-supervised learning from images with a joint-embedding predictive architecture,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 15619–15629. 12

    2023

  26. [26]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning, 2020, pp. 1597–1607

    2020

  27. [27]

    Contrastive learning of subject-invariant EEG representations for cross-subject emotion recognition,

    X. Shen, X. Liu, X. Hu, D. Zhang, and S. Song, “Contrastive learning of subject-invariant EEG representations for cross-subject emotion recognition,” IEEE Trans. Affect. Comput., vol. 14, no. 3, pp. 2496– 2511, 2022

    2022

  28. [28]

    Multi-Scale Hyperbolic Contrastive Learning for Cross-Subject EEG Emotion Recognition,

    J. Chang, Z. Zhang, Y. Qian, and P. Lin, “Multi-Scale Hyperbolic Contrastive Learning for Cross-Subject EEG Emotion Recognition,” IEEE Trans. Affect. Comput., 2025

    2025

  29. [29]

    Multi-Scale Masked Autoencoders for Cross-Session Emotion Recognition,

    M. Pang, H. Wang, J. Huang, C.-M. Vong, Z. Zeng, and C. Chen, “Multi-Scale Masked Autoencoders for Cross-Session Emotion Recognition,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2024. [ 3 0 ] V. J . L a w he r n , A . J . S o l o n, N . R . W a yt o w i c h, S . M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: a compact convolutio...

    2024

  30. [30]

    EEG conformer: Convolutional transformer for EEG decoding and visualization,

    Y. Song, Q. Zheng, B. Liu, and X. Gao, “EEG conformer: Convolutional transformer for EEG decoding and visualization,” IEEE Transactions on Neural Systems and Rehabilitation Engineering , vol. 31, pp. 710–719, 2022

    2022

  31. [31]

    Instance-adaptive graph for EEG emotion recognition,

    T. Song, S. Liu, W. Zheng, Y. Zong, and Z. Cui, “Instance-adaptive graph for EEG emotion recognition,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2020, pp. 2701–2708

    2020

  32. [32]

    PGCN: Pyramidal graph convolutional network for EEG emotion recognition,

    M. Jin, C. Du, H. He, T. Cai, and J. Li, “PGCN: Pyramidal graph convolutional network for EEG emotion recognition,” IEEE Trans. Multimedia, 2024

    2024

  33. [33]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy et al. , “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  34. [34]

    Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning,

    Y. Zhang, B. Hooi, D. Hu, J. Liang, and J. Feng, “Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning,” Adv. Neural Inf. Process. Syst. , vol. 34, pp. 29848– 29860, 2021

    2021

  35. [35]

    A large finer-grained affective computing EEG dataset,

    J. Chen, X. Wang, C. Huang, X. Hu, X. Shen, and D. Zhang, “A large finer-grained affective computing EEG dataset,” Sci. Data, vol. 10, no. 1, p. 740, 2023

    2023

  36. [36]

    EmotionMeter: A Multimodal Framework for Recognizing Human Emotions,

    W.-L. Zheng, W. Liu, Y. Lu, B.-L. Lu, and A. Cichocki, “EmotionMeter: A Multimodal Framework for Recognizing Human Emotions,” IEEE Trans. Cybern., vol. 49, no. 3, pp. 1110–1122, 2019, doi: 10.1109/TCYB.2018.2797176

    doi:10.1109/tcyb.2018.2797176 2019

  37. [37]

    Comparing Recognition Performance and Robustness of Multimodal Deep Learning Models for Multimodal Emotion Recognition,

    W. Liu, J.-L. Qiu, W.-L. Zheng, and B.-L. Lu, “Comparing Recognition Performance and Robustness of Multimodal Deep Learning Models for Multimodal Emotion Recognition,” IEEE Trans. Cogn. Dev. Syst. , vol. 14, pp. 715–729, 2022

    2022

  38. [38]

    Seed-vii: A multimodal dataset of six basic emotions with continuous labels for emotion recognition,

    W.-B. Jiang, X.-H. Liu, W.-L. Zheng, and B.-L. Lu, “Seed-vii: A multimodal dataset of six basic emotions with continuous labels for emotion recognition,” IEEE Trans. Affect. Comput., 2024

    2024

  39. [39]

    Temporal Aware Mixed Attention-based Convolution and Transformer Network (MACTN) for EEG Emotion Recognition,

    X. Si, D. Huang, Y. Sun, and D. Ming, “Temporal Aware Mixed Attention-based Convolution and Transformer Network (MACTN) for EEG Emotion Recognition,” arXiv preprint arXiv:2305.18234, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Identifying stable patterns over time for emotion recognition from EEG,

    W.-L. Zheng, J.-Y. Zhu, and B.-L. Lu, “Identifying stable patterns over time for emotion recognition from EEG,” IEEE Trans. Affect. Comput., vol. 10, no. 3, pp. 417–429, 2017

    2017

  41. [41]

    Fusing frequency-domain features and brain connectivity features for cross-subject emotion recognition,

    C. Chen, Z. Li, F. Wan, L. Xu, A. Bezerianos, and H. Wang, “Fusing frequency-domain features and brain connectivity features for cross-subject emotion recognition,” IEEE Trans. Instrum. Meas., vol. 71, pp. 1–15, 2022

    2022

  42. [42]

    HSA-Former: Hierarchical Spatial Aggregation Transformer for EEG-Based Emotion Recognition,

    J. Huang et al. , “HSA-Former: Hierarchical Spatial Aggregation Transformer for EEG-Based Emotion Recognition,” IEEE Trans. Comput. Soc. Syst., 2025

    2025

  43. [43]

    Emotion recognition using multi-modal data and machine learning techniques: A tutorial and review,

    J. Zhang, Z. Yin, P. Chen, and S. Nichele, “Emotion recognition using multi-modal data and machine learning techniques: A tutorial and review,” Information Fusion, vol. 59, pp. 103–126, 2020

    2020

  44. [44]

    Emt: A novel transformer for generalized cross-subject eeg emotion recognition,

    Y. Ding et al. , “Emt: A novel transformer for generalized cross-subject eeg emotion recognition,” IEEE Trans. Neural Netw. Learn. Syst., 2025

    2025

  45. [45]

    Bipartite graph adversarial network for subject-independent emotion recognition,

    M. Niaki, S. Y. Dharia, Y. Chen, and C. E. Valderrama, “Bipartite graph adversarial network for subject-independent emotion recognition,” IEEE J. Biomed. Health Inform., 2025

    2025