Pith. sign in

REVIEW 4 major objections 6 minor 33 references

PeakFlow claims that EEG-based dynamic affective trajectory prediction is best done with explicit peak-aware refinement: a coarse-to-refined model that preserves global trends while correcting peak timing, peak value, and terminal-peak bias

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 14:40 UTC pith:3WVABHC6

load-bearing objection Useful, well-engineered fix for terminal-peak bias in EEG emotion trajectory prediction, but the headline peak metrics are partly trained onto the same labels they measure — claims need recalibration. the 4 major comments →

arxiv 2607.18671 v1 pith:3WVABHC6 submitted 2026-07-21 cs.HC

PeakFlow: Peak-Guided Coarse-to-Refined Modeling for EEG-Based Dynamic Affective Trajectory Prediction

Hao Tang , Songyun Xie , Xinzhou Xie , Can Liao , Xin Zhang , Bohan Li , Zhongyu Tian , Dalu Zheng This is my paper
classification cs.HC
keywords EEG emotion recognitiondynamic affective trajectorypeak detectionterminal-peak biasmasked temporal modelingvector quantizationresidual calibrationcontinuous emotion intensity
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.

The paper argues that predicting a continuous emotion intensity curve from EEG should not stop at global curve fitting, because the emotionally salient peak—when intensity is maximum—is what people remember and act on. It identifies a structural failure in masked temporal trajectory models: the predicted peak is systematically shifted toward the end of a trial. The authors propose PeakFlow, a three-stage framework that first learns discrete EEG temporal tokens, then predicts a coarse affective flow with a masked Transformer, and finally applies a lightweight residual refiner trained with a peak-centered objective. On leave-one-subject-out evaluation on SEED-VII, PeakFlow reduces the false-terminal peak rate from 62.56% to 5.42% and improves peak-time and peak-value errors while also slightly improving global MSE and PCC. The paper also presents auxiliary results on FIRMED suggesting the refined peak intensities preserve ordinal emotion levels.

Core claim

The central discovery is that peak-centered temporal structure can be recovered from EEG without sacrificing global trajectory accuracy, by decoupling coarse affective-flow modeling from a bounded residual calibration module. Terminal-peak bias—where the predicted maximum intensity lands in the last 10% of the trial even though the ground-truth peak is earlier—afflicts strong dynamic baselines (62.56% false-terminal rate on SEED-VII), yet a peak-guided refiner that sees only trajectory-level cues (coarse intensity, normalized time position, distance to the terminal region, local temporal difference) and is trained with a peak-zone-weighted, terminal-suppressed objective cuts this to 5.42%. T

What carries the argument

The load-bearing mechanism is a three-stage coarse-to-refined pipeline. Stage I is a feature-level vector-quantized tokenizer that discretizes EEG window features into 64 codebook states, providing auxiliary discrete targets. Stage II is a masked dynamic Transformer that predicts a coarse intensity trajectory from continuous EEG features while also predicting the discrete token indices. Stage III is the peak-guided bounded residual refiner: a small temporal convolutional network that takes trajectory-aware cues from the coarse prediction, estimates a soft peak-probability sequence, and outputs a bounded residual correction (1 + η·q_peak) ⊙ α·tanh(ρ), added to the coarse flow and clipped to [

Load-bearing premise

The whole pipeline, including training labels and peak-centered metrics, assumes SEED-VII's dense continuous intensity annotations are reliable ground truth and that each trial has a single well-defined argmax peak; noisy annotations, inter-rater disagreement, or frequent plateau peaks would undermine both the supervision and the evaluation.

What would settle it

Re-annotate a subset of SEED-VII trials with multiple raters marking the affective peak location and intensity; if inter-rater agreement is low, or if a large fraction of ground-truth trajectories have flat plateaus instead of isolated maxima, then the single-argmax peak definition and the false-terminal-peak metric lose their validity. Alternatively, retrain PeakFlow without the terminal penalty but with all other training components intact and measure the false-terminal rate; if it rises substantially above 5.42%, the headline FTR result is driven by the penalty suppressing terminal peaks ra

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Point-wise regression metrics like MSE and PCC are insufficient for dynamic affective prediction; peak-time error, peak-value error, and false-terminal peak rate should be reported alongside them.
  • The peak-guided refiner can be attached to any frozen coarse trajectory predictor, since it consumes only trajectory cues, not raw EEG features.
  • Applications that depend on the most emotionally salient moment—multimedia highlight detection, neurofeedback, adaptive user interfaces—would inherit the improved peak localization.
  • The peak-end rule from retrospective emotion psychology becomes directly testable in EEG-based models, because predicted peaks and endings can now be localized reliably.
  • Auxiliary results on FIRMED suggest the same refinement transfers to sparse event-level ordinal intensity labels, broadening its use beyond dense trajectory datasets.

Where Pith is reading between the lines

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

  • The dramatic false-terminal-rate drop may be partly an artifact of the terminal penalty suppressing all elevated terminal predictions; a fair test would compare models trained with and without the penalty under identical peak-time evaluation, which the paper's ablations only partially isolate.
  • Because the ground-truth peak is defined by a single argmax, trials with flat or plateau peaks are poorly served; a peak-mass or multi-peak formulation might be more robust, and PeakFlow's ranking could change on datasets with plateau-dominant annotations.
  • The refiner's trajectory-level cues make the calibration stage dataset- and sensor-agnostic, so the same post-hoc calibration could be applied to facial-expression, EDA, or heart-rate trajectories without retraining the base model.
  • The false-terminal rate depends on the arbitrary 10% terminal-region definition; comparing FTR across papers that use different r_term values is not meaningful, and a fixed r_term may penalize models differently on trials of different lengths.

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

4 major / 6 minor

Summary. The paper proposes PeakFlow, a three-stage framework for EEG-based dynamic affective trajectory prediction. Stage I learns a discrete EEG temporal tokenizer via vector quantization; Stage II trains a masked Transformer to predict a coarse affective intensity trajectory; Stage III adds a lightweight TCN-based residual refiner that uses trajectory-aware cues and a peak-centered calibration objective (trajectory fitting, peak-zone emphasis, peak-probability localization, terminal suppression, residual regularization) to correct peak timing, peak value, and false-terminal peaks. The method is evaluated with leave-one-subject-out experiments on SEED-VII against window-wise and temporal-sequence baselines, reporting improved global metrics (MSE, MAE, PCC, R²) and peak-centered metrics (normalized peak-time error, peak-value error, false-terminal peak rate). An auxiliary evaluation on FIRMED examines ordinal intensity consistency under sparse event-level annotations. The central claim is that PeakFlow improves both global trajectory fitting and peak-centered temporal reliability.

Significance. If the empirical claims hold, the paper addresses a real limitation of global point-wise regression for affective trajectory prediction: terminal-peak bias and inaccurate peak localization. The staged coarse-to-refined architecture is well motivated, and the ablation study supports the contribution of the peak-centered objective, terminal penalty, bounded residual design, and temporal cues. The paper also ships code and introduces arguably useful peak-centered evaluation metrics. However, the significance is currently conditional: the headline peak metrics are directly optimized by the training losses, all results are point estimates without variance or significance testing, and the single-argmax peak label is not validated for stability or inter-rater reliability. These issues must be resolved before the reported improvements can be taken as evidence of a generalizable modeling advantage.

major comments (4)
  1. [§IV-B, Eqs. (47)–(52); §III-E.3, Eqs. (41)–(45)] The peak-centered metrics are not independent of the training objective. FTR (Eq. 52) defines the terminal region using r_term=0.10, the same r_term=0.10 used in L_end (Eq. 45), which explicitly penalizes positive prediction error in that terminal region. Likewise, E_time and E_value are computed from the single argmax peak (Eq. 47), while L_pz (Eq. 42) up-weights fitting around q_zone (Eq. 30) centered at the same annotated peak p_i, and L_prob (Eq. 44) trains q_peak to match q_zone. The FTR drop from 62.56% to 5.42% therefore partly reflects optimization of the evaluation metric itself, not an independent measure of temporal localization. Please report robustness of the peak-centered results to: (i) alternative r_term values (e.g., 0.05, 0.15, 0.20); (ii) tolerance-based peak-time error not coinciding with the training zone (e.g., |p̂-p|>d); and (iii) peaks redefined after temporal smo
  2. [§IV-C, Table I] All reported results are single point estimates with no variance, confidence intervals, or significance tests. The LOSO protocol yields 20 subjects, but no per-subject statistics are reported. The global improvements are small (MSE 0.0758→0.0733, PCC 0.5523→0.5754) and may be within intersubject or run-to-run variability. Please report per-subject means and standard deviations, the number of random seeds, and paired significance tests across subjects (e.g., Wilcoxon signed-rank) for MSE, MAE, PCC, peak-time error, peak-value error, and FTR. This is necessary to support the claim that PeakFlow improves over EEGDancer and other baselines.
  3. [§III-A, Eq. (4); §IV-A] The peak-centered formulation rests entirely on the single-argmax definition p_i = argmax_t y_i,t. The paper provides no evidence that this label is stable or reliable: no inter-rater agreement for SEED-VII annotations, no characterization of isolated maxima versus plateaus, and no analysis of near-ties. If a substantial fraction of trials have flat or noisy maxima, the argmax position is an unstable target for both training and evaluation. Please report: (i) the distribution of peak margins (max y_i minus second-highest valid y_i) and the fraction of trials with near-ties/plateaus; (ii) inter-rater agreement on peak location if available; and (iii) robustness of Table I conclusions when peaks are defined with a plateau-aware rule (e.g., midpoint of the maximal plateau) or after light temporal smoothing.
  4. [§III-E.1, Eq. (30); Algorithm 1; §IV-D] The peak-zone radius R, which defines q_zone and therefore controls L_pz and L_prob, is never given a default value anywhere in the paper. Algorithm 1 lists R as a required input, but the implementation details and hyperparameter sensitivity analysis (Fig. 7) do not state R or vary it. Because L_pz directly up-weights the region around the annotated peak, the reported peak-time and peak-value improvements may depend on this unstated hyperparameter. Please specify the default R and include a sensitivity analysis over R.
minor comments (6)
  1. [§III-D, Eq. (21)] Equation (21) appears to have a typographical error: b_r appears both inside the GELU argument and outside it. One of these is likely unintended.
  2. [§III-E.3, Eq. (39)] L_Δ is used in Eq. (39) but never explicitly defined. Please define the first-order temporal-difference loss.
  3. [§IV-E, Eq. (53)] The FIRMED auxiliary evaluation does not specify the size of the event-centered window W_i,k, nor whether the models were retrained on FIRMED or transferred from SEED-VII. Please clarify the protocol and report results with variance across subjects/sessions.
  4. [References] References [6] and [17] are duplicates, as are [7] and [18]. Reference [24] contains an unresolved placeholder ('please verify volume, issue, and page numbers'). Please clean up the bibliography.
  5. [§V-C] The limitations paragraph in §V-C appropriately lists offline prediction, single-peak modeling, and unimodal signals, but it does not mention the dependence of all peak-centered results on the reliability of the single-argmax annotation. This should be added in light of the concerns above.
  6. [Fig. 7] The hyperparameter curves in Fig. 7 are shown as single lines without error bars or repeated-run variation. Given the small global-metric differences (e.g., PCC differences of 0.01–0.02), error bars or multiple runs are needed to interpret the stability claim.

Circularity Check

0 steps flagged

No significant circularity: the peak-centered metrics align with the Stage-III training losses, but the main results are evaluated under held-out LOSO folds on an external benchmark, so the improvements are not logically forced.

full rationale

PeakFlow is an empirical supervised method, not a first-principles derivation. The Stage-III objective (Eqs. 38-46) does share design elements with the evaluation metrics: L_pz upweights the region around the labeled peak p_i (Eqs. 30, 42-43), L_prob supervises a peak-probability map with the same zone labels (Eqs. 30, 44), and L_end penalizes positive terminal-region errors using r_term=0.10 (Eq. 45), the same terminal fraction used in FTR (Eqs. 49-52). This means the peak-centered metrics are not adversarially independent of the training loss. However, there is no identity between loss and metric: L_end is a smooth squared penalty on terminal overestimation, while FTR is an argmax event (Eq. 51), and L_pz is a weighted regression error, not the absolute argmax distance E_time. More importantly, the model is trained on LOSO training folds and evaluated on held-out subjects, so the reported improvements reflect generalization of a learned function rather than a fitted constant being renamed as a prediction. No parameter is fitted directly to test peaks, and the paper does not use the test labels in training. The claimed derivation chain therefore does not reduce to its inputs. The FIRMED evidence is cited from the authors' own prior annotation paper (Ref. [25]), but it is explicitly auxiliary and complementary to the main SEED-VII results, so this self-citation is not load-bearing for the central claim. The stability and annotation reliability of the single-argmax peak definition is a validity concern, but it is not circularity: even a noisy label can be predicted under held-out evaluation. Overall, the core empirical claim is self-contained against an external benchmark and is not circular by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central contribution is an empirical method; it introduces no new physical entities. The key free parameters are the loss weights and geometric thresholds that jointly define the peak-centered objective and the evaluation metrics. The most consequential axiom is that the SEED-VII single-peak annotation is reliable, since it drives both training and evaluation.

free parameters (5)
  • Terminal region proportion r_term = 0.10
    Defines the terminal region in both the training loss L_end (Eq. 45) and the FTR metric (Eq. 52); hand-chosen, not derived.
  • Peak-zone radius R
    Defines the binary peak-zone label q_zone (Eq. 30) used for peak-probability supervision and peak-zone weighting; value not reported in text.
  • Stage-III loss weights (λ_peak, λ_end, λ_res) and ω_Δ, ω_prob, ω_pz = defaults selected on validation set (exact values not listed)
    Balances trajectory, peak, end, and residual terms in Eq. 38; chosen by validation, not by theory.
  • Residual scale α and peak-gate strength η
    Control magnitude and gating of the bounded residual correction in Eqs. 35-36; values not stated in paper.
  • VQ codebook size K and latent dimension d_z = K=64, d_z=64
    Selected on validation set; governs the capacity of the discrete tokenizer.
axioms (4)
  • domain assumption SEED-VII continuous affective intensity annotations are valid, reliable ground truth for affective trajectories.
    The entire training/evaluation rests on these annotations (Section IV-A). The paper does not assess inter-rater reliability or noise.
  • domain assumption The affective peak, defined as the single argmax of the intensity trajectory (Eq. 4), is a well-defined, meaningful temporal event.
    Used to construct peak labels, losses, and metrics; multi-peak or plateau structures are ignored (acknowledged in Section V-C).
  • ad hoc to paper r_term = 0.10 is an appropriate definition of the terminal region for both the terminal-suppression loss and the FTR metric.
    This parameter appears symmetrically in training (Eq. 45) and evaluation (Eq. 52), making the FTR improvement partly dependent on this choice.
  • ad hoc to paper The residual refiner form (g_i = 1 + η q_peak, r_corr = α g_i ⊙ tanh(ρ_i)) is a sufficient inductive bias for bounded peak correction.
    This particular gating design is motivated heuristically; no theoretical guarantee that it preserves global trend.

pith-pipeline@v1.3.0-alltime-deepseek · 18607 in / 11647 out tokens · 125936 ms · 2026-08-01T14:40:14.827636+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of PeakFlow: Peak-Guided Coarse-to-Refined Modeling for EEG-Based Dynamic Affective Trajectory Prediction." pith.science (2026). https://pith.science/paper/3WVABHC6

@misc{pith2026260718671,
  author       = {Pith},
  title        = {Pith review of: PeakFlow: Peak-Guided Coarse-to-Refined Modeling for EEG-Based Dynamic Affective Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WVABHC6}},
  note         = {Machine review of arXiv:2607.18671}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Most existing EEG-based emotion recognition studies formulate affective decoding as static category prediction, although emotions elicited by continuous stimulation evolve over time, accumulate, reach peak intensity, and then recover. This motivates EEG-based dynamic affective trajectory prediction, which estimates continuous affective intensity curves from sequential EEG observations. Existing temporal regression models can capture coarse intensity trends but often fail to preserve peak-centered structure, leading to inaccurate peak timing and terminal-peak bias, where the predicted maximum is shifted toward the end of a trial. To address this issue, we propose PeakFlow, a peak-guided coarse-to-refined framework for EEG-based dynamic affective trajectory prediction. PeakFlow first learns a coarse affective flow through EEG temporal tokenization and masked temporal modeling, then applies a lightweight residual refiner for peak-guided bounded calibration. The refiner uses trajectory-aware cues and a peak-centered objective combining global trajectory consistency, peak-zone emphasis, peak-probability localization, terminal suppression, and residual regularization. This design preserves the global affective trend while correcting peak misalignment, peak-value deviation, and false-terminal predictions. Leave-one-subject-out experiments on SEED-VII show that PeakFlow improves both global trajectory fitting and peak-centered temporal reliability over strong dynamic modeling baselines. Auxiliary evaluation on FIRMED further suggests its potential for sparse peak-centered ordinal intensity analysis. These results highlight the importance of peak-aware modeling for temporally faithful EEG-based dynamic emotion prediction. Code is available at https://github.com/jukebox333/PeakFlow.

Figures

Figures reproduced from arXiv: 2607.18671 by Bohan Li, Can Liao, Dalu Zheng, Hao Tang, Songyun Xie, Xin Zhang, Xinzhou Xie, Zhongyu Tian.

Figure 1
Figure 1. Figure 1: Evidence of terminal-peak bias in SEED-VII. For each trial, the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Stage-wise framework of PeakFlow. Stage I learns a feature-level EEG tokenizer with a learnable codebook. Stage II freezes the tokenizer and trains a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the learned EEG VQ latent states on SEED-VII. (A) The VQ codebook embeddings are projected into a two-dimensional space using [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Across different subjects and emotion categories, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Peak-centered analysis and terminal-peak bias correction on SEED-VII. (A) Distribution of normalized peak positions. (B) False-terminal peak rate. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples of peak-aware trajectory refinement on SEED-VII. Each subplot shows the ground-truth affective intensity trajectory, the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Subject-wise and emotion-wise consistency analysis on SEED-VII. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity analysis of PeakFlow with respect to the Stage-III loss weights. We vary the peak weight [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Auxiliary ordinal intensity-level visualization on FIRMED. The [PITH_FULL_IMAGE:figures/full_fig_p012_8.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

33 extracted references · 2 linked inside Pith

  1. [1]

    R. W. Picard,Affective Computing. Cambridge, MA, USA: MIT Press, 1997

    1997

  2. [2]

    Deap: A database for emotion analysis using physiological signals,

    S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “Deap: A database for emotion analysis using physiological signals,”IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18–31, 2012

    2012

  3. [3]

    Differential entropy feature for eeg-based emotion classification,

    R.-N. Duan, J.-Y . Zhu, and B.-L. Lu, “Differential entropy feature for eeg-based emotion classification,” inProceedings of the 6th Interna- tional IEEE/EMBS Conference on Neural Engineering. IEEE, 2013, pp. 81–84

    2013

  4. [4]

    Investigating critical frequency bands and channels for eeg-based emotion recognition with deep neural networks,

    W.-L. Zheng and B.-L. Lu, “Investigating critical frequency bands and channels for eeg-based emotion recognition with deep neural networks,” IEEE Transactions on Autonomous Mental Development, vol. 7, no. 3, pp. 162–175, 2015

    2015

  5. [5]

    Eegnet: A compact convolutional neural network for eeg-based brain–computer interfaces,

    V . J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “Eegnet: A compact convolutional neural network for eeg-based brain–computer interfaces,”Journal of Neural Engineering, vol. 15, no. 5, p. 056013, 2018

    2018

  6. [6]

    Eeg emotion recognition using dynamical graph convolutional neural networks,

    T. Song, W. Zheng, P. Song, and Z. Cui, “Eeg emotion recognition using dynamical graph convolutional neural networks,”IEEE Transactions on Affective Computing, vol. 11, no. 3, pp. 532–541, 2020

    2020

  7. [7]

    Eeg-based emotion recognition using regularized graph neural networks,

    P. Zhong, D. Wang, and C. Miao, “Eeg-based emotion recognition using regularized graph neural networks,”IEEE Transactions on Affective Computing, vol. 13, no. 3, pp. 1290–1301, 2022

    2022

  8. [8]

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

    W.-B. Jianget al., “Seed-vii: A multimodal dataset of six basic emotions with continuous labels for emotion recognition,”IEEE Transactions on Affective Computing, 2024, early Access

    2024

  9. [9]

    Eegdancer: Dynamic emotion latent space masked modeling with reinforce- ment learning for eeg continuous emotion prediction,

    Z. Zhou, W. Ye, L. Zhang, G. Huang, and Z. Liang, “Eegdancer: Dynamic emotion latent space masked modeling with reinforce- ment learning for eeg continuous emotion prediction,”arXiv preprint arXiv:2606.05855, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, 2017, pp. 5998–6008

    2017

  11. [11]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems, 2017, pp. 6306–6315

    2017

  12. [12]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inPro- ceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, 2019, pp. 4171–4186

    2019

  13. [13]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 000–16 009

    2022

  14. [14]

    Duration neglect in retrospective evaluations of affective episodes,

    B. L. Fredrickson and D. Kahneman, “Duration neglect in retrospective evaluations of affective episodes,”Journal of Personality and Social Psychology, vol. 65, no. 1, pp. 45–55, 1993

    1993

  15. [15]

    Patients’ memories of painful medical treatments: Real-time and retrospective evaluations of two minimally invasive procedures,

    D. A. Redelmeier and D. Kahneman, “Patients’ memories of painful medical treatments: Real-time and retrospective evaluations of two minimally invasive procedures,”Pain, vol. 66, no. 1, pp. 3–8, 1996

    1996

  16. [16]

    Eeg-based emotion recognition using frequency domain features and support vector machines,

    X.-W. Wang, D. Nie, and B.-L. Lu, “Eeg-based emotion recognition using frequency domain features and support vector machines,” in Neural Information Processing, ser. Lecture Notes in Computer Science, vol. 7062. Springer, 2011, pp. 734–743

    2011

  17. [17]

    Eeg emotion recognition using dynamical graph convolutional neural networks,

    T. Song, W. Zheng, P. Song, and Z. Cui, “Eeg emotion recognition using dynamical graph convolutional neural networks,”IEEE Transactions on Affective Computing, vol. 11, no. 3, pp. 532–541, 2020, early access in 2018

    2020

  18. [18]

    Eeg-based emotion recognition using regularized graph neural networks,

    P. Zhong, D. Wang, and C. Miao, “Eeg-based emotion recognition using regularized graph neural networks,”IEEE Transactions on Affective Computing, vol. 13, no. 3, pp. 1290–1301, 2022, early access in 2020

    2022

  19. [19]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  20. [20]

    Learning transferable features with deep adaptation networks,

    M. Long, Y . Cao, J. Wang, and M. I. Jordan, “Learning transferable features with deep adaptation networks,” inProceedings of the 32nd International Conference on Machine Learning, 2015, pp. 97–105

    2015

  21. [21]

    Deep coral: Correlation alignment for deep domain adaptation,

    B. Sun and K. Saenko, “Deep coral: Correlation alignment for deep domain adaptation,” inEuropean Conference on Computer Vision Work- shops. Springer, 2016, pp. 443–450

    2016

  22. [22]

    Intensity profiles of emotional experience over time,

    P. Verduyn, I. Van Mechelen, F. Tuerlinckx, K. Meers, and H. Van Coil- lie, “Intensity profiles of emotional experience over time,”Cognition and Emotion, vol. 23, no. 7, pp. 1427–1443, 2009

    2009

  23. [23]

    Emotion dynamics,

    P. Kuppens and P. Verduyn, “Emotion dynamics,”Current Opinion in Psychology, vol. 17, pp. 22–26, 2017

    2017

  24. [24]

    Mgeed: A multimodal genuine emotion and expression detection database,

    Y . Wanget al., “Mgeed: A multimodal genuine emotion and expression detection database,”IEEE Transactions on Affective Computing, 2023, please verify volume, issue, and page numbers before final submission

    2023

  25. [25]

    From coarse to fine-grained emotion annotation: An immediate recall paradigm with validation through physiological evidence and recognition performance,

    H. Tang, S. Xie, X. Xie, C. Liao, X. Zhang, B. Li, Z. Tian, and D. Zheng, “From coarse to fine-grained emotion annotation: An immediate recall paradigm with validation through physiological evidence and recognition performance,”arXiv preprint arXiv:2507.02350, 2025

    arXiv 2025

  26. [26]

    E- time: Emotion trend inspired multi-task sparse mask neural network for multimodal emotion recognition,

    S. Zhang, J. Wang, Z. Feng, X. Ning, X. Jin, Z. Jia, and Y . Lin, “E- time: Emotion trend inspired multi-task sparse mask neural network for multimodal emotion recognition,”IEEE Journal of Biomedical and Health Informatics, 2026

    2026

  27. [27]

    Do emotions last longer than moods?

    P. Verduyn, E. Delvaux, H. Van Coillie, F. Tuerlinckx, and I. Van Meche- len, “Do emotions last longer than moods?”Cognition and Emotion, vol. 23, no. 6, pp. 1237–1255, 2009

    2009

  28. [28]

    Determinants of the shape of emotion intensity profiles,

    P. Verduyn, I. Van Mechelen, and E. Frederix, “Determinants of the shape of emotion intensity profiles,”Cognition and Emotion, vol. 26, no. 8, pp. 1486–1495, 2012

    2012

  29. [29]

    Experiences extended across time: Evalua- tion of moments and episodes,

    C. Varey and D. Kahneman, “Experiences extended across time: Evalua- tion of moments and episodes,”Journal of Behavioral Decision Making, vol. 5, no. 3, pp. 169–185, 1992

    1992

  30. [30]

    When more pain is preferred to less: Adding a better end,

    D. Kahneman, B. L. Fredrickson, C. A. Schreiber, and D. A. Redelmeier, “When more pain is preferred to less: Adding a better end,”Psycholog- ical Science, vol. 4, no. 6, pp. 401–405, 1993

    1993

  31. [31]

    Gestalt characteristics of experiences: The defining features of summarized events,

    D. Ariely and Z. Carmon, “Gestalt characteristics of experiences: The defining features of summarized events,”Journal of Behavioral Decision Making, vol. 13, no. 2, pp. 191–201, 2000

    2000

  32. [32]

    All’s well that ends (and peaks) well? a meta-analysis of the peak-end rule and duration neglect,

    B. Alaybek, R. S. Dalal, S. Fyffe, J. A. Aitken, Y . Zhou, X. Qu, A. Roman, and J. I. Baines, “All’s well that ends (and peaks) well? a meta-analysis of the peak-end rule and duration neglect,”Organizational Behavior and Human Decision Processes, vol. 170, p. 104149, 2022

    2022

  33. [33]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,”arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018