Pith. sign in

REVIEW 5 minor 33 references

Reported EEG emotion-recognition accuracy is a property of the evaluation pipeline, not the classifier: the same saved runs score 88.9% with test-set checkpoint selection and 78.6% without, and drop to 53.5% on unseen participants.

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 · deepseek-v4-flash

2026-08-01 03:54 UTC pith:5XMQCHOA

load-bearing objection Careful, honest, narrow methods paper; the same-trajectory checkpoint-selection contrast is the real contribution and it holds up.

arxiv 2607.27655 v1 pith:5XMQCHOA submitted 2026-07-30 cs.LG

Evaluation Protocols and Cross-Subject Generalization in EEG Emotion Recognition

Hanting Suo , Yuwen Li This is my paper
classification cs.LG
keywords EEG emotion recognitionaffective computingevaluation protocolcheckpoint selectiontest-set reusebenchmark reproducibilitysubject-disjoint generalizationsubject-level inference
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.

This paper argues that a reported EEG emotion-recognition accuracy number does not describe the model but the entire evaluation procedure wrapped around it, and that subject-dependent, subject-disjoint, and cross-session results answer different questions and must be reported separately. The controlled demonstration uses one archived graph-convolutional network on two public datasets: on 30 identical saved training trajectories, choosing the checkpoint by peeking at the test set raised mean window accuracy from 0.7855 to 0.8892, a paired gain of 0.1036; under strict five-fold subject-disjoint evaluation with no target-data access, the same pipeline fitted its training participants to 0.9990 yet scored 0.5348 on entirely held-out participants. That gap rules out a broken optimizer as the explanation for poor cross-subject results, without isolating subject identity from preprocessing, representation, or implementation factors. The paper's prescription is a three-layer protocol record — target quantity, development procedure, reporting rule — and a ten-point checklist, so that benchmarks stop treating 'accuracy on SEED' as one number. Why it matters: if accurate, much of the spread between published EEG emotion-recognition results is protocol variation, and protocol-blind leaderboards are comparing unlike quantities.

Core claim

The central claim: an EEG accuracy score describes a complete evaluation pipeline, not a classifier, and no two scores are comparable until the pipeline's target, development procedure, and reporting rule are separated. On 30 saved SEED trajectories of one archived graph-convolutional network, checkpoints selected by repeated test-set evaluation reached 0.8892 accuracy versus 0.7855 at the fixed terminal epoch (paired difference 0.1036). Under five-fold subject-disjoint evaluation without target-data access, the same runs scored 0.9990 on training participants but 0.5348 on held-out participants (95% BCa [0.4667, 0.5985]); the SEED-IV estimate, 0.3954, is secondary because its protocol-match

What carries the argument

The load-bearing object is a three-layer protocol record: a target-estimand specification (data, outcome, prediction unit, aggregation, metric, population weighting); a development-and-estimation procedure (split construction, admissible target information, selection rule, training design); and a reporting specification (resampling unit, interval type, endpoint status). The argument's engine is a selection inequality: the maximum of K noisy test-set scores has non-negative expected optimism over the best candidate's expected accuracy, so choosing a checkpoint by repeatedly scoring the test set inflates reported performance. A design-effect approximation, D = 1 + (m−1)ρ, carries the conceptua

Load-bearing premise

The measured gaps (0.1036 selection effect; roughly 0.46 train-to-held-out-subject gap) rest on one archived implementation of the graph-convolutional network being a faithful representative of the method — if that code path, including its historical all-ones Chebyshev filter, behaves unlike other implementations, or the public SEED-IV reference used a different split or preprocessing, the numbers could be shifted by software rather than protocol.

What would settle it

Score the 30 saved SEED trajectories against a session-3 test split never opened during training or selection: if test-selected checkpoints no longer beat epoch-80 checkpoints, the 0.1036 gap is selection optimism; if the edge persists, it is not. Run the strict five-fold subject-disjoint protocol with an independently written implementation of the same network: if held-out accuracy moves far from 0.5348, the gap belongs to the archived code, not the protocol. Recover the historical SEED-IV artifacts and test whether the reference 0.5239 matches within two points; a match would resolve the one

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

If this is right

  • Subject-dependent, subject-disjoint, and cross-session accuracies should be published as answers to different questions; a headline number without a protocol record (target, access, selection, scoring unit, weighting) supports no specific deployment claim.
  • Every use of test-set labels that can influence selection — checkpoint choice, early stopping, rerunning — must be counted; the matched contrast shows the practice can add 0.1036 mean window accuracy on the same saved trajectories.
  • Low held-out-subject accuracy should not be read as training failure: here training-participant accuracy exceeded 0.99 while unseen participants scored near 0.53, so training fit must be reported alongside generalization.
  • Participant-level rankings were not stable enough to justify persistent weighting of low-performing subjects: median split-half Spearman was 0.1700 and cross-session 0.0470 on the primary dataset, and the ordering depended on feature representation.
  • Development-stage success does not license final claims: a tail-risk ensemble that passed its development criteria did not establish a positive accuracy gain over the comparator on held-out trials in the separate final evaluation.

Where Pith is reading between the lines

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

  • My inference: if a 0.10 selection effect and a roughly 0.46 subject-boundary gap appear on one archived pipeline, published EEG leaderboards are likely reshuffled by the same protocol coordinates; any benchmark number should be treated as protocol-relative until target access and checkpoint selection are disclosed.
  • My inference: near-perfect training fit with near-chance held-out accuracy suggests the learned graph structure is person-specific; a testable extension is to measure how much of the gap closes when learned adjacency matrices are reused across subjects or target-batch statistics are admitted.
  • My inference: the paper's checklist could be operationalized as a machine-readable protocol field attached to every reported accuracy, so downstream comparisons are filterable by protocol equality — a natural next step beyond the paper.
  • My inference: the 0.1036 matched-trajectory contrast is a direct price for adaptive test-set reuse; the same saved-trajectory design could measure the cost of hyperparameter tuning, feature selection, and early stopping in neighbouring EEG and physiological-classification settings.

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

0 major / 5 minor

Summary. The paper argues that reported accuracy in EEG emotion recognition is a property of the complete evaluation procedure—target estimand, development procedure, and reporting rule—not of the classifier alone. It formalizes this with a three-layer protocol record (I, V, Q), gives conditions for score comparability, proves an elementary selection inequality showing that selecting checkpoints by repeated test-set evaluation has non-negative expected optimism, and proposes a minimum reporting checklist. Empirically, using one archived DGCNN pathway on SEED and SEED-IV, it shows that test-set-selected checkpoints outperform terminal-epoch checkpoints by 0.1036 on the same 30 saved trajectories; that strict five-fold subject-disjoint evaluation yields held-out-subject accuracy around 0.5348 on SEED while training-subject accuracy is 0.9990; and that supporting analyses on participant-rank stability and a tail-risk ensemble do not support persistent low-performing-subject weighting or a transferred recognition gain. The paper is explicit that the specific magnitudes are illustrative and implementation-dependent, and that only the paired checkpoint contrast is a controlled decomposition.

Significance. The contribution is methodological rather than architectural. The paper's central claim—that EEG accuracy scores are not comparable from dataset and metric names alone—is well supported. The paired checkpoint-selection contrast in §5.1 is a strong piece of evidence because it holds the saved trajectories, model family, and epoch schedule fixed, isolating the checkpoint-selection rule. The selection inequality in §3.3 is elementary but correct, and the paper carefully avoids overinterpreting it as a universal bias correction. The subject-disjoint results are appropriately qualified as implementation- and protocol-conditional, and the unresolved SEED-IV compatibility difference is honestly reported rather than concealed. The paper ships frozen artifacts, audit scripts, and machine-readable configurations, which strengthens reproducibility. If the field adopts the proposed protocol record, this could meaningfully improve comparability of EEG emotion-recognition studies.

minor comments (5)
  1. [§3.1] There is a typographical/formatting issue: 'T arget-free subject-disjointdenotes' should be 'Target-free subject-disjoint denotes'. Also, the later text prefers the phrase 'subject-disjoint evaluation without target-data access'; using one consistent term would reduce ambiguity.
  2. [§4.3] The phrase 'historical all-ones Chebyshev T0 implementation' is cryptic to readers outside the graph-convolution literature. A one-sentence explanation of what this means and why it matters for the archived pathway would improve accessibility.
  3. [§4.1] The competition four-cell analysis is described as 'pre-specified before its results were examined,' but no evidence of the pre-specification (e.g., a dated document or a hash of a specification file) is provided. Given the paper's emphasis on auditable protocols, stating how pre-specification was recorded would be consistent with the paper's own standards.
  4. [§5.2] In Fig. 4, the captions say 'Dotted vertical lines mark the median validation-selected epoch' and the figure text shows 'median selected epoch: 16' and '10'. It may be clearer to state in the caption that these are medians over the 15 dataset-fold-seed cells, as is implied in the text.
  5. [§6.5] The limitations paragraph is thorough. It may be worth adding one sentence in §6.5 or the conclusion noting that the conditional BCa intervals do not account for the model-selection process across folds, since this is already stated in §4.6 but could be more prominent.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical measurement study whose protocol-sensitivity findings are observed contrasts, not fitted predictions or self-citation chains.

full rationale

The paper's central claim—that reported EEG emotion-recognition accuracy depends on the complete evaluation procedure—is supported by direct measurements, not by a derivation that assumes the conclusion. The headline checkpoint-selection contrast (0.8892 vs 0.7855 on the same 30 saved trajectories) is read off fixed artifacts: the selection inequality in §3.3 only establishes the non-negative expected optimism of test-set-based selection and is explicitly not used to compute the realized 0.1036 difference. The subject-disjoint results are likewise observed outcomes under a fixed 9/3/3 split and frozen seeds, with the paper itself repeatedly stating that the train-to-held-out gap does not isolate subject identity. No parameter is fitted to a target and then renamed as a prediction, and no load-bearing self-citation or imported uniqueness theorem appears: the methodological novelty is an evaluation checklist and an empirical stress test, not a result forced by prior work of the same authors. The unresolved SEED-IV compatibility difference is disclosed as a limitation rather than smoothed over. There is no circular step satisfying the requirement of a quoted equation or definition that reduces to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or ontological entities. Its formal three-layer protocol record is a methodological taxonomy, and the ledger's free parameters are the hand-set experimental decisions that define the protocol comparisons. The empirical claims are conditional measurements rather than fitted derivations.

free parameters (5)
  • Terminal epoch 80 baseline = epoch 80
    The paired checkpoint-selection contrast compares test-set-selected checkpoints against this fixed terminal epoch on the saved trajectories; a different fixed epoch would change the reported 0.1036 gap.
  • Validation checkpoint-selection rule = max pooled validation-window macro-F1; earliest-epoch tie
    All strict subject-disjoint estimates depend on this hand-chosen development criterion; another selection rule would select different checkpoints and change the estimates.
  • Subject-disjoint fold partition seed = random.Random(2024); 9/3/3 subject split
    The strict five-fold split is one realized partition; the subject-disjoint estimates and BCa intervals are conditional on this partition and would vary with other realizations.
  • Compatibility tolerance = ±2 percentage points
    The pre-specified two-percentage-point tolerance determines whether the SEED-IV compatibility difference is classified as unresolved; a wider tolerance would change that classification.
  • Ranking-reliability engineering gates = 0.30 split-half; 0.40 cross-session
    These pre-specified thresholds determine the conclusion that participant rankings are not stable enough for persistent weighting; different gates would alter the interpretation.
axioms (5)
  • domain assumption The archived LibEER DGCNN implementation plus DE/LDS one-second representation is a faithful implementation of DGCNN for SEED/SEED-IV.
    All numerical results are produced on this pipeline; audits passed for the local path, but exact equivalence to historical public runs, especially SEED-IV, could not be established.
  • domain assumption The public reference values (SEED 0.8255; SEED-IV 0.5239) correspond to the same ordered subject-dependent split as the local reconstruction.
    The protocol-matched reference check is meaningful only if the historical split and preprocessing align with the local run; SEED-IV historical artifacts were unavailable, leaving the difference unresolved.
  • domain assumption Subject-level BCa intervals conditional on fitted fold models and fixed partitions are appropriate uncertainty summaries for the conditional risk.
    The paper explicitly states that these intervals are not procedure-level intervals and do not integrate over new splits, retraining, or repeated selection; this is a disclosed interpretive choice.
  • domain assumption The 30 matched subject-session trajectories are genuinely the same saved training runs, differing only in checkpoint selection.
    The checkpoint-selection contrast is controlled only if reloading saved checkpoints and re-evaluating them does not introduce hidden state changes; the paper provides checksums and audit scripts but not an independent third-party verification.
  • standard math E[max_k A_k] ≥ max_k E[A_k] for noisy test-accuracy candidates.
    Used in §3.3 to show that selecting the maximum noisy test score has non-negative expected optimism; this is elementary and correct.

pith-pipeline@v1.3.0-daily-deepseek · 23656 in / 15642 out tokens · 166471 ms · 2026-08-01T03:54:01.203281+00:00 · methodology

0 comments
read the original abstract

Reported accuracy in electroencephalography (EEG) emotion recognition depends on the complete evaluation procedure, not only the classifier. We separate the target quantity, development procedure, and reporting rule, then use one archived dynamical graph convolutional neural network (DGCNN) pathway on SEED and SEED-IV as an illustrative case. In a protocol-matched subject-dependent check, the SEED result was within 1.47 percentage points of the public reference value; the 3.40-point SEED-IV difference remained unresolved. Across 30 matched SEED subject-session trajectories, checkpoint selection based on repeated test-set evaluation increased mean window accuracy from 0.7855 at epoch 80 to 0.8892. Under five-fold subject-disjoint evaluation, validation-selected checkpoints achieved training-participant trial accuracies of 0.9990 on SEED and 0.9920 on SEED-IV. Accuracy for entirely held-out participants was 0.5348 (95% conditional subject-level bias-corrected and accelerated [BCa] interval [0.4667, 0.5985]) on SEED. The SEED-IV estimate was 0.3954 ([0.3343, 0.4648]) and is reported only as secondary sensitivity evidence because its protocol-matched compatibility check remained unresolved. The observed train-to-held-out-subject gaps are inconsistent with simple optimization underfitting, but they do not isolate subject identity from implementation, preprocessing, representation, or distributional factors. Supporting analyses further showed that participant rankings depended on representation and time scale, while a development-selected tail-risk ensemble did not establish a positive gain in a separate final evaluation. Subject-dependent, subject-disjoint, and cross-session results should therefore be reported as answers to different questions.

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

33 extracted references · 6 canonical work pages

  1. [1]

    Proceedings of the National Academy of Sciences 99(10):6562–6566

    Ambroise C, McLachlan GJ (2002) Selection bias in gene extraction on the basis of microarray gene-expression data. Proceedings of the National Academy of Sciences 99(10):6562–6566. https://doi.org/10.1073/pnas.102102699

    doi:10.1073/pnas.102102699 2002

  2. [2]

    https://doi.org/10.1080/01621459.2023.2197686

    Bates S, Hastie T, Tibshirani R (2024) Cross-validation: What does it estimate and how well does it do it? Journal of the American Statistical Association 119(546):1434–1445. https://doi.org/10.1080/01621459.2023.2197686

    arXiv 2024

  3. [3]

    Journal of Machine Learning Research 5:1089–1105

    Bengio Y, Grandvalet Y (2004) No unbiased estimator of the variance of k- fold cross-validation. Journal of Machine Learning Research 5:1089–1105. URL https://www.jmlr.org/papers/v5/grandvalet04a.html

    2004

  4. [4]

    URL https://arxiv.org/abs/24 06.10366, arXiv:2406.10366

    Binette O, Reiter JP (2024) Improving the validity and practical usefulness of AI/ML evaluations using an estimands framework. URL https://arxiv.org/abs/24 06.10366, arXiv:2406.10366

    Pith/arXiv arXiv 2024

  5. [5]

    In: Proceedings of Machine Learning and Systems, pp 747–769, URL https://proceedings.mlsys.org/paper files/paper/2021/hash/0 184b0cd3cfb185989f858a1d9f5c1eb-Abstract.html

    Bouthillier X, Delaunay P, Bronzi M, et al (2021) Accounting for variance in machine learning benchmarks. In: Proceedings of Machine Learning and Systems, pp 747–769, URL https://proceedings.mlsys.org/paper files/paper/2021/hash/0 184b0cd3cfb185989f858a1d9f5c1eb-Abstract.html

    2021

  6. [6]

    Frontiers in Neuroscience 18:1373515

    Brookshire G, Kasper J, Blauch NM, et al (2024) Data leakage in deep learning studies of translational EEG. Frontiers in Neuroscience 18:1373515. https://doi.or g/10.3389/fnins.2024.1373515

    arXiv 2024

  7. [7]

    Journal of Machine Learning Research 11:2079–2107

    Cawley GC, Talbot NLC (2010) On over-fitting in model selection and subsequent selection bias in performance evaluation. Journal of Machine Learning Research 11:2079–2107. URL https://jmlr.org/papers/v11/cawley10a.html

    2010

  8. [8]

    Computers in Biology and Medicine 196:110608

    Del Pup F, Zanola A, Tshimanga LF, et al (2025) The role of data partitioning on the performance of EEG-based deep learning models in supervised cross-subject analysis: A preliminary study. Computers in Biology and Medicine 196:110608. https://doi.org/10.1016/j.compbiomed.2025.110608

    arXiv 2025

  9. [9]

    In: 2013 6th International IEEE/EMBS Conference on Neural Engineering (NER), pp 81–84, https://doi.org/10.1109/NER.2013.6695876

    Duan RN, Zhu JY, Lu BL (2013) Differential entropy feature for EEG-based emotion classification. In: 2013 6th International IEEE/EMBS Conference on Neural Engineering (NER), pp 81–84, https://doi.org/10.1109/NER.2013.6695876

    arXiv 2013

  10. [10]

    Multimedia Tools and Applications 83:79647–79694

    Erat K, Sahin EB, Dogan F, et al (2024) Emotion recognition with EEG-based brain–computer interfaces: A systematic literature review. Multimedia Tools and Applications 83:79647–79694. https://doi.org/10.1007/s11042-024-18259-z

    work page doi:10.1007/s11042-024-18259-z 2024

  11. [11]

    Journal of Machine Learning Research 17(59):1–35

    Ganin Y, Ustinova E, Ajakan H, et al (2016) Domain-adversarial training of neural networks. Journal of Machine Learning Research 17(59):1–35. URL https: //jmlr.org/papers/v17/15-239.html 30

    2016

  12. [12]

    https://doi.org/10.1016/j.bspc.2026.110020

    Gil D, Hern´ andez-Sabat´ e A (2026) What is the across-subject performance of your artificial intelligence method for EEG analysis? Biomedical Signal Processing and Control 120:110020. https://doi.org/10.1016/j.bspc.2026.110020

    arXiv 2026

  13. [13]

    Neuroscience Bulletin 41(3):434–448

    Guan Z, Zhang X, Huang W, et al (2025) A method for detecting depression in adolescence based on an affective brain–computer interface and resting-state EEG signals. Neuroscience Bulletin 41(3):434–448. https://doi.org/10.1007/s12264-024 -01319-7, published online 20 November 2024

    work page doi:10.1007/s12264-024 2025

  14. [14]

    IEEE Transactions on Affective Computing 14(2):998–1011

    Huang W, Wu W, Lucas MV, et al (2023) Neurofeedback training with an electroencephalogram-based brain–computer interface enhances emotion regulation. IEEE Transactions on Affective Computing 14(2):998–1011. https://doi.org/10.1 109/TAFFC.2021.3134183

    arXiv 2023

  15. [15]

    Patterns 4:100804

    Kapoor S, Narayanan A (2023) Leakage and the reproducibility crisis in machine- learning-based science. Patterns 4:100804. https://doi.org/10.1016/j.patter.2023. 100804

    doi:10.1016/j.patter.2023 2023

  16. [16]

    URL https: //arxiv.org/abs/2505.18175, arXiv:2505.18175

    Kukhilava N, Tsmindashvili T, Kalandadze R, et al (2025) Evaluation in EEG emotion recognition: State-of-the-art review and unified framework. URL https: //arxiv.org/abs/2505.18175, arXiv:2505.18175

    Pith/arXiv arXiv 2025

  17. [17]

    In: Proceedings of the 1st Workshop on 4D Micro-Expression Recognition (4DMR 2025), pp 69–77, URL https://ceur-ws.or g/Vol-4115/paper7.pdf

    Lei P, Wu M, Yi W, et al (2025) Impact of trial-wise and test data leakage on EEG-based emotion classification. In: Proceedings of the 1st Workshop on 4D Micro-Expression Recognition (4DMR 2025), pp 69–77, URL https://ceur-ws.or g/Vol-4115/paper7.pdf

    2025

  18. [18]

    Frontiers in Neuroscience 15:611653

    Li J, Li S, Pan J, et al (2021) Cross-subject EEG emotion recognition with self-organized graph neural network. Frontiers in Neuroscience 15:611653. https: //doi.org/10.3389/fnins.2021.611653

    arXiv 2021

  19. [19]

    Frontiers in Human Neuroscience 17:1169949

    Li J, Pan W, Huang H, et al (2023) STGATE: Spatial-temporal graph attention network with a transformer encoder for EEG-based emotion recognition. Frontiers in Human Neuroscience 17:1169949. https://doi.org/10.3389/fnhum.2023.1169949

    arXiv 2023

  20. [20]

    Frontiers in Computational Neuroscience 20:1865513

    Li Z, Wu X, Hao Y, et al (2026) Cross-subject generalization for EEG emotion recognition: A review of methods, challenges, and future trends. Frontiers in Computational Neuroscience 20:1865513. https://doi.org/10.3389/fncom.2026.186 5513

    work page doi:10.3389/fncom.2026.186 2026

  21. [21]

    IEEE Transactions on Affective Computing 16(4):3596–3613

    Liu H, Yang S, Zhang Y, et al (2025) LibEER: A comprehensive benchmark and algorithm library for EEG-based emotion recognition. IEEE Transactions on Affective Computing 16(4):3596–3613. https://doi.org/10.1109/TAFFC.2025.360 5833

    work page doi:10.1109/taffc.2025.360 2025

  22. [22]

    American Sociological Review 86(3):532–565

    Lundberg I, Johnson R, Stewart BM (2021) What is your estimand? defining the target quantity connects statistical evidence to theory. American Sociological Review 86(3):532–565. https://doi.org/10.1177/00031224211004187 31

    doi:10.1177/00031224211004187 2021

  23. [23]

    Scientific Reports 16:18646

    Margaret MJ, Banu NMM, Madhumithaa S, et al (2026) On the prediction models for brain signal-based emotion recognition. Scientific Reports 16:18646. https://doi.org/10.1038/s41598-026-47622-3

    work page doi:10.1038/s41598-026-47622-3 2026

  24. [24]

    Journal of Machine Learning Research 22(164):1–20

    Pineau J, Vincent-Lamarre P, Sinha K, et al (2021) Improving reproducibility in machine learning research: A report from the NeurIPS 2019 reproducibility program. Journal of Machine Learning Research 22(164):1–20. URL https://jmlr .org/papers/v22/20-303.html

    2021

  25. [25]

    Ecography 40(8):913–929

    Roberts DR, Bahn V, Ciuti S, et al (2017) Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8):913–929. https://doi.org/10.1111/ecog.02881

    doi:10.1111/ecog.02881 2017

  26. [26]

    In: International Conference on Learning Representations, URL https://openreview.net/forum?id=ryxGuJrFvS

    Sagawa S, Koh PW, Hashimoto TB, et al (2020) Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=ryxGuJrFvS

    2020

  27. [27]

    IEEE Transactions on Affective Computing 11(3):532–541

    Song T, Zheng W, Song P, et al (2020) EEG emotion recognition using dynamical graph convolutional neural networks. IEEE Transactions on Affective Computing 11(3):532–541. https://doi.org/10.1109/TAFFC.2018.2817622

    arXiv 2020

  28. [28]

    In: Proceedings of the 38th International Conference on Machine Learning, pp 9913–9922, URL https://proceedings.mlr.press/v139/summers21a.ht ml

    Summers C, Dinneen MJ (2021) Nondeterminism and instability in neural network optimization. In: Proceedings of the 38th International Conference on Machine Learning, pp 9913–9922, URL https://proceedings.mlr.press/v139/summers21a.ht ml

    2021

  29. [29]

    In: Computer Vision – ECCV 2016 Workshops, pp 443–450, https: //doi.org/10.1007/978-3-319-49409-8 35

    Sun B, Saenko K (2016) Deep CORAL: Correlation alignment for deep domain adaptation. In: Computer Vision – ECCV 2016 Workshops, pp 443–450, https: //doi.org/10.1007/978-3-319-49409-8 35

    work page doi:10.1007/978-3-319-49409-8 2016

  30. [30]

    BMC Bioinformatics 7:91

    Varma S, Simon R (2006) Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics 7:91. https://doi.org/10.1186/1471-2105-7-91

    doi:10.1186/1471-2105-7-91 2006

  31. [31]

    NeuroImage 180:68–77

    Varoquaux G (2018) Cross-validation failure: Small sample sizes lead to large error bars. NeuroImage 180:68–77. https://doi.org/10.1016/j.neuroimage.2017.06.061

    doi:10.1016/j.neuroimage.2017.06.061 2018

  32. [32]

    IEEE Transactions on Autonomous Mental Development 7(3):162–175

    Zheng WL, Lu BL (2015) Investigating critical frequency bands and channels for EEG-based emotion recognition with deep neural networks. IEEE Transactions on Autonomous Mental Development 7(3):162–175. https://doi.org/10.1109/TA MD.2015.2431497

    arXiv 2015

  33. [33]

    scale", class weight=

    Zheng WL, Liu W, Lu Y, et al (2019) EmotionMeter: A multimodal framework for recognizing human emotions. IEEE Transactions on Cybernetics 49(3):1110–1122. https://doi.org/10.1109/TCYB.2018.2797176 32 Supplementary Information for: Evaluation Protocols and Cross-Subject Generalization in EEG Emotion Recognition Hanting Suo1,2 and Yuwen Li1,2* 1School of In...

    arXiv 2019