- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
EEG decoders generalize poorly to unseen subjects. MGT weights source data by estimated transferability at two levels. A reference model is first trained on source and target-calibration data and then frozen. In its feature space, a subject-level weight comes from entropy-regularized optimal transport between each source subject's class-wise features and the target's (Eq. 2-3), and a sample-level weight from distances to target class prototypes and a cluster-to-target OT cost (Eq. 4-8). The two are multiplied, normalized per class and clipped (Eq. 9-10), kept fixed, and ramped in over the first 100 epochs (Eq. 11) on the source terms of the GAT objective (Eq. 12). Inference needs neither source data nor target calibration labels (Sec. 2.3).
Evaluation covers seven datasets and five paradigms (motor imagery, emotion, imagined speech, RSVP, ERN), with subject-averaged balanced accuracy, its subject-wise standard deviation, Kappa and AUROC (Sec. 3.1). MGT reports the best primary metric on all six datasets in the main text, with margins over re-run GAT (official updated) from +0.20 on ERN to +2.35 on RSVP (Tables 2 and 7).
Critical strengths
S1. The method is simple to add. Transfer weights are computed once in a frozen reference space and stay fixed during final training (Sec. 2.1, 2.3), and they only replace the source average in the loss (Eq. 12). In principle the approach could be added to backbones other than GAT, although only GAT is tested.
S2. For six of the seven datasets, preprocessing is described carefully enough to rule out the leakage that is easy to introduce in cross-subject EEG work. App. B.1 states that normalization statistics come only from the training or calibration partition (on BCI-IV-2a, "Per-channel z-score statistics are estimated exclusively from the training partition"). The SEED paragraph gives no normalization procedure.
Critical weaknesses
W1. The gaps to the literature baselines in Table 2 cannot be attributed to MGT. MGT and GAT (official updated) use a substantial labeled target-calibration set (80% of the target subject on RSVP, Sessions 1-4 with 240 trials on ERN, and the target's full official training partition on BCI-IV-2a; App. B.1), while App. C.1 says the other results are "literature-reported" and never states their calibration budget. The 18.0-point gap over BioCodec on ERN (78.89 vs 60.90) and the 9.8-point gap over CST-TVA-DRTL on RSVP (83.70 vs 73.95) are already present for GAT (official updated) before MGT is applied. Either rerun the baselines under this protocol or restrict quantitative claims to the GAT comparison.
W2. Hyperparameters appear to have been selected on the evaluation data, and the per-subject evidence for the coarse-grained factor is weak. App. B.1 defines only a calibration and an evaluation partition for each dataset, and the word validation does not appear. Figure 6(a) reports balanced accuracy on BCI-IV-2a and peaks at lambda_proto = 0.5, the default used for all reported results; with no validation split, the only held-out data are the target evaluation partition. Using the per-subject values in App. D.2-D.4, a two-sided paired sign test does not find the removal of Coarse-Grained Transfer significant on any dataset: the full model wins on 11 of 15 subjects on SEED (p = 0.12), 6 of 8 decided on BCI-IV-2a (p = 0.29) and 3 of 5 decided on NM000252 (p = 1.00). A Wilcoxon signed-rank test reaches p = 0.03 on SEED but not on BCI-IV-2a (p = 0.08) or NM000252 (p = 1.00), and with five decided pairs the sign test cannot reach 0.05 on NM000252 at all. By the sign test, removing Progressive Transfer on SEED (p = 0.002) and Fine-Grained Transfer on BCI-IV-2a (p = 0.039) is significant. The conclusion's statement that the ablations support all three components should be qualified for the coarse-grained one.
W3. The related work is too narrow. MEKT (Zhang & Wu, 2020) already estimates subject-level transferability to choose source subjects for a target. It appears in the paper only as the MEKT-R baseline on RSVP and is not discussed with the selective-transfer work, so what the coarse-grained weight adds over it is not addressed.
W1 and W2 can be addressed in revision: W1 by rerunning the baselines under the same calibration protocol or narrowing the claims to the GAT comparison, and W2 by stating how the hyperparameters were chosen and qualifying the ablation conclusion.
Questions for authors
Q1. How many runs and seeds are behind Tables 2, 3 and 7, and which data were used to select lambda_proto, E_ramp, m, beta_min, tau_t and the clipping bounds, and to stop training early in the Figure 6 runs? The Std columns are defined across subjects (Sec. 3.1), and four of seven datasets show gains under one point over GAT (official updated) against subject-wise standard deviations of 6.00 to 13.22.
Q2. Table 3 has no row with both granularities removed. Setting all w_i = 1 reduces Eq. (12) to the ordinary source average, so GAT (official updated) should be that reference, but the paper does not say whether it shares the seed and schedule of the ablation runs. Could you add that row, with paired per-subject tests for each ablation?
Q3. How does MGT relate to existing work on weighting sources? Besides MEKT, source instances have been reweighted using labeled target data, for example in TrAdaBoost (Dai et al., 2007), and multi-source adaptation has learned per-domain weights, for example Domain Aggregation Networks (Wen et al., ICML 2020). In its discussion of selective transfer, the introduction cites Wang et al. (2024b) and Cui et al. (2025). Is the claim that the two levels had not been combined before, or that these particular estimators are new?
Q4. In Table 2 the Std column is exactly 0.00 for xDAWN-SVM on RSVP (11 subjects) and for CBraMod and STELAR-base on ERN (16 subjects), and the other ERN literature baselines lie between 0.20 and 0.89, whereas GAT (official updated) and MGT show 10.80 and 10.27 on the same data. These values do not look like the subject-wise standard deviation defined in Sec. 3.1. What does the Std column measure for each literature baseline, and under what calibration budget were these values obtained?
Q5. For every new target subject, MGT trains an 80-epoch reference model on the source pool together with that subject's calibration data, estimates the transfer weights, and then trains the final model from a new initialization (Sec. 2.1; 80 epochs in Sec. 3.1). The conclusion notes this added computation, but no training time is reported. How long does this take per target compared with GAT alone? Together with the labeled calibration set it requires (App. B.1), this decides how practical the method is for a new user.
6 comments
Response to Reviewer p33B(Part 2)
W2. Hyperparameter Selection and Statistical Evidence for Coarse-Grained Transfer
Response: Thank you for raising these important questions. We clarify the target-data partitioning and model selection protocol, as well as the contribution of coarse-grained transfer. Target data partitioning and model selection. The Target Calibration Set and Target Test Set are two separate, non-overlapping partitions of the target subject's data. The labeled calibration set is used for adaptation, transferability estimation, and model selection, while the test set is reserved for final evaluation. Specifically, BCI-IV-2a uses the official training and test partitions, RSVP adopts a stratified 80%/20% split, and ERN uses Sessions 1–4 for calibration and Session 5 for testing. These protocols are described in Section 3.1 and Appendix B.1.
As stated in Appendix B.2, “The key hyperparameters were selected using a simple grid search based on performance on the target calibration set.” The selected configuration is summarized in Table 6. The nominal cluster size (m) is fixed at 64 across datasets. Checkpoints are selected according to calibration performance, as described in Section 3.1, with the separate test set used for final evaluation.
For the sensitivity experiments in Figure 6, each target subject has a maximum training budget of 200 epochs, with early stopping patience set to 30 epochs. Training terminates when the monitored calibration performance fails to improve for 30 consecutive epochs. Early stopping is applied independently to each target subject.
Contribution of coarse-grained transfer. As shown in Table 4, removing coarse-grained transfer reduces the average BAcc across the five reported datasets, demonstrating its contribution to the overall framework. The subject-wise results in Appendix D.2–D.4 further reveal that its contribution varies across target subjects.
Regarding the reviewer's statistical analysis, the Wilcoxon signed-rank test yields (p=0.03) on SEED, while the corresponding results are (p=0.08) on BCI-IV-2a and (p=1.00) on NM000252. The sign-test results are (p=0.29), (p=0.12), and (p=1.00), respectively. These results reflect differences in how the two tests evaluate paired subject-level improvements, with the Wilcoxon test additionally accounting for the magnitude and ranking of paired differences.
We further distinguish these component-level analyses from the evaluation of the complete MGT framework. As reported in Section 3.2, Table 3, MGT improves BAcc over GAT (official updated) by 2.02 percentage points across the nine target subjects, with improvements on seven subjects and a Holm-adjusted (p=0.0391). This provides additional statistical evidence for the effectiveness of the complete MGT framework.
Taken together, the average ablation results support the contribution of coarse-grained transfer, while the subject-wise analyses characterize its varying effects across target subjects. The statistical comparison with GAT further demonstrates the effectiveness of the integrated MGT framework.
Response to Reviewer p33B(Part 3)
W3. Related Work on Selective Transfer Methods
Response: Thank you for suggesting these relevant studies. We have expanded Appendix A (Related Work) to discuss existing selective transfer methods and clarify the methodological contributions of MGT. The revised manuscript introduces three complementary directions:
1.MEKT (Zhang & Wu, 2020): “aligns EEG covariance representations on a Riemannian manifold and performs domain adaptation in the tangent space for unsupervised cross-subject classification.”
2.TrAdaBoost (Dai et al., 2007): “iteratively adjusts source-instance contributions within a boosting framework to exploit labeled target data.”
3.Domain Aggregation Networks (Wen et al., 2020): “dynamically balance contributions from multiple source domains to reduce negative transfer.”
These discussions are included in Appendix A, together with the corresponding references.
We further describe the specific design of MGT in the same section: “MGT jointly estimates subject-level transferability through class-conditional optimal transport and trial-level transferability through prototype and local-cluster costs in a frozen reference space.”
“The resulting transfer factors are fused and progressively incorporated into source-dependent training objectives, capturing complementary variations across source subjects and individual trials.”
The methodological formulation is detailed in Sections 2.1–2.3, Eqs. (1)–(12). Specifically, the coarse-grained factor characterizes source-subject compatibility, while the fine-grained factor evaluates individual trial compatibility using target-class prototypes and local cluster structures. Their multiplicative fusion and progressive integration provide a unified mechanism for regulating source contributions during cross-subject EEG adaptation.
Q1. Experimental Repetitions, Hyperparameter Selection, and Stopping Criteria
Response: Thank you for requesting these clarifications.Regarding experimental repetitions, MGT and GAT (official updated) on BCI-IV-2a were evaluated using five random seeds. MGT was also evaluated using five seeds (2022–2026) on RSVP and ERN. The corresponding results are reported in Appendix D.6 and D.9–D.10. The other MGT benchmark results use single runs, while literature baselines retain the results reported in their original studies.
Five-seed evaluation summary across three EEG datasets.
| Dataset | Method | BAcc (Mean ± SD) (%) | Additional Metric |
|---|---|---|---|
| BCI-IV-2a | MGT | 83.11 ± 0.08 | -- |
| BCI-IV-2a | GAT (official updated) | 81.09 ± 0.28 | -- |
| RSVP | MGT | 86.11 ± 1.43 | AUROC: 91.67 ± 0.25 |
| ERN | MGT | 79.14 ± 0.20 | AUROC: 83.95 ± 0.03 |
For the statistical comparisons in Table 3, each target subject is treated as one paired observation. The five random seeds are averaged at the subject level rather than treated as independent observations, as clarified in Appendix D.7. The BCI-IV-2b comparison, originally presented in Table 7, is now included in Appendix D.1.
Regarding hyperparameter selection, the Target Calibration Set and Target Test Set are separate data partitions. Hyperparameters are selected through grid search based on target calibration performance, as described in Appendix B.2. The final configurations are summarized in Table 6. The nominal cluster size m is fixed at 64 across datasets. Figure 6 further examines the sensitivity to the prototype–cluster balance and progressive ramp duration on BCI-IV-2a.
For the Figure 6 experiments, each target subject has a maximum training budget of 200 epochs. Early stopping is applied independently to each subject, with a patience of 30 epochs. Training terminates when no performance improvement is observed for 30 consecutive epochs.
Checkpoint selection is based on target calibration performance, while the separate Target Test Set is reserved for final evaluation, as described in Section 3.1.
Response to Reviewer p33B(Part 1)
We sincerely thank the reviewer for the thoughtful and constructive comments. We appreciate the positive assessment of the simplicity of MGT and its evaluation across diverse EEG paradigms. We have carefully considered the concerns regarding comparison fairness, model selection, statistical evidence, related work, and computational efficiency. In the following, we address each point in detail and describe the corresponding revisions.
W1. Attribution of Improvements over Literature Baselines
Response: Thank you for raising this important concern. We agree that performance differences between methods using different target-calibration protocols cannot be attributed solely to MGT. The original GAT study (Song et al., 2023) provides relevant evidence for the contribution of source knowledge and domain adaptation. On BCI-IV-2a, the target-only model achieved 62.96% accuracy, compared with 64.54% for source–target training without adaptation and 76.58% for full GAT. On BCI-IV-2b, full GAT achieved 84.44% accuracy, compared with 79.95% for the target-only model. These results are summarized in Appendix C.2, Table 8 of our revised manuscript. They demonstrate the importance of source–target adaptation in the original GAT framework and help explain why differences in training protocols can substantially affect cross-method comparisons.
To ensure a clearer comparison, we distinguish literature-reported results from controlled experiments. On BCI-IV-2a, seven comparison methods were re-implemented using the same target calibration/test partition as MGT. For the remaining datasets, literature-reported results are retained as references rather than treated as strictly controlled comparisons. Our primary method comparison focuses on GAT (official updated), which uses the same v0.72 backbone and evaluation protocol as MGT. These distinctions are clarified in Appendix C.1–C.2.
Under the shared evaluation settings, MGT improves BAcc over GAT (official updated) by 2.02, 0.45, 1.60, 0.84, 2.41, and 0.25 percentage points on BCI-IV-2a, OpenBMI, SEED, NM000252, RSVP, and ERN, respectively. An additional improvement of 0.55 percentage points is observed on BCI-IV-2b (Appendix D.1).
We have also added subject-wise statistical analyses on BCI-IV-2a to further assess the significance of these improvements. The Friedman test identifies significant differences among the nine compared methods (χ²(8) = 48.80, p < 0.001). Two-sided paired Wilcoxon signed-rank tests with Holm correction show significant improvements over all eight comparison methods (all adjusted p ≤ 0.0391). In particular, MGT outperforms GAT (official updated) on seven of nine target subjects, with a mean paired improvement of 2.02 percentage points (adjusted p = 0.0391). These results are reported in Section 3.2, Table 3, with subject-wise data and five-seed results provided in Appendix D.6–D.7.
We have therefore clarified the scope of our quantitative claims. The literature-reported results provide broader context, while the controlled comparisons with GAT (official updated) and the supplementary statistical analyses provide more direct evidence of MGT's effectiveness.