acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Learning to Adapt Spatial–Spectral Relations for Cross-Dataset EEG Emotion Recognition

18 views

Abstract

Cross-dataset EEG emotion recognition is challenging because spatial–spectral relations that are useful for one subject or recording condition may not transfer to another. We show that even within predictive spatial relation groups, frequency- specific relations can have different effects, and useful adjustments vary across targets. We therefore propose TRACE (Target-conditioned Relation Adaptation for Cross-dataset EEG), which learns rules for adjusting spatial–spectral rela- tions from labeled source domains and applies these rules to each unlabeled tar- get. TRACE represents channel–band interactions with shared relation types and adapts target-specific adjustments to their relative weights from a domain repre- sentation aggregated from unlabeled target trials. It learns these rules from source- domain adjustments validated by lower classification loss on separate source trials and constrains target coefficients to a source-supported region defined by source- domain predictions. TRACE projects the difference between graph representa- tions before and after relation adjustments into a residual that refines the base EEG representation learned directly from the input features for emotion classification. Across eight transfer directions among four public EEG emotion datasets, TRACE achieves the highest accuracy among the compared methods in every direction. Target-substitution experiments further show that adjustments inferred for the cur- rent target outperform those transferred from other targets, while TRACE remains robust to electrode masking and feature noise. These results support learning how relations adapt rather than enforcing a fixed structure across domains.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG emotion recognition, cross-dataset generalization, unsupervised domain adaptation, graph neural networks, spatial-spectral relations, transfer learning, target-conditioned adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uESJOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge of transferring spatial–spectral relations across datasets for EEG emotion recognition and proposes TRACE, which aims to learn how to adaptively adjust relation weights according to the characteristics of different target domains. The method learns a relation adaptation rule from labeled source domains, infers target-specific adjustment coefficients from unlabeled target-domain data, and incorporates the resulting relational changes into EEG representations for emotion classification. TRACE achieves improved performance across 8 transfer directions involving 4 EEG emotion datasets.

Critical strengths
  1. Novel perspective on relation adaptation. Rather than transferring a fixed spatial–spectral structure across datasets, TRACE learns a transferable rule for adjusting relation weights according to target-domain context, offering a clear perspective on cross-dataset EEG emotion recognition.
  2. Target-specific adaptation. TRACE infers target-specific relation adjustment coefficients from unlabeled target-domain context and constrains them to a source-supported region, limiting unconstrained extrapolation during adaptation.
  3. Consistent empirical performance across evaluated transfer directions. Experiments across four public EEG emotion datasets and eight transfer directions show that TRACE achieves the highest reported accuracy and Macro-F1 among the compared methods in the evaluated settings. Ablation, target-substitution, and input-perturbation experiments provide additional evidence supporting the proposed design.
Critical weaknesses
  1. Dependence on the availability of unlabeled target-domain data. TRACE operates in a transductive setting and uses all available unlabeled windows from the target domain for adaptation. This assumption may limit its applicability in scenarios where only a small amount of target data is available or data arrive sequentially. The robustness of the method under limited target-data availability has not been sufficiently investigated.
  2. Concerns about statistical reliability. Although TRACE achieved performance improvements in different migration directions, the standard deviation of performance is particularly large in many settings. Meanwhile, the results are conducted under three random seeds, which may not be sufficient to prove the stability and significance of the model.
  3. The motivation shown in Figure 1 is difficult to understand. Figure 1 contains several key diagnostic experiments, but it is not easy to understand them clearly. It would be better if the description could be improved.
Questions for authors
  1. Adaptation with limited target data: How does TRACE perform when only a fraction of the unlabeled target-domain data is available? It would be better to report results under different target-data budgets, and discuss the feasibility of adapting to streaming target data.
  2. Statistical reliability: Could you provide results across additional random seeds, report variability across runs separately from variation across target recordings, and conduct paired statistical significance tests against the strongest baselines?
  3. Contribution of relationship modeling: To what extent does graph based relationship modeling itself bring benefits compared to target condition relationship adjustment?
6 comments
AuthorsOpenReview ↗

## Response to Reviewer uESJ — Part 1/3

Thank you for the constructive feedback and for recognizing the value of target-specific relation adaptation. We conducted supplementary evaluations of target-data availability and random-seed variability. We also clarified the motivation in Figure 1 and the interpretation of the graph-modeling ablations. Revisions addressing this review are marked in dark green.

W1 and Q1. Adaptation with limited unlabeled target data and streaming targets.

We evaluated TRACE under four target-data budgets on SEED → SEED-IV, using 10 fixed target recordings and three random seeds. Each SEED-IV recording contains 24 complete trials. We constructed nested, label-agnostic subsets containing 6, 12, 18, and 24 trials, corresponding to 25%, 50%, 75%, and 100% of the available trials.

The budget restriction applies to the entire target-side procedure, rather than to context estimation alone. At each budget, target min–max normalization, context construction, relation-coefficient inference, clustering, pseudo-label generation, and downstream adaptation use the selected unlabeled trials. The source-trained model is shared across the four budgets within each seed, and each budget completes the same adaptation procedure. All conditions are evaluated on the same complete target recordings, including the trials previously used for unlabeled adaptation. Target labels are used for evaluation, not adaptation.

Available target trialsAccuracy (%)Macro-F1 (%)BAcc (%)
25% (6/24)53.91 ± 7.2751.64 ± 7.4353.23 ± 7.71
50% (12/24)54.33 ± 4.1952.94 ± 4.7054.32 ± 5.14
75% (18/24)54.64 ± 4.1452.83 ± 3.4854.01 ± 3.01
100% (24/24)55.03 ± 4.6852.96 ± 4.6355.97 ± 5.25

Entries report mean ± sample SD across three seed-wise recording averages. With 25% rather than 100% of the trials, accuracy decreases by 1.12 percentage points, macro-F1 by 1.32 points, and BAcc by 2.74 points. BAcc exhibits the largest decrease. These results characterize TRACE's behavior under reduced target-data availability in the evaluated offline transductive setting.

The model design provides several reasons for the observed behavior.

First, TRACE learns the context-to-relation adaptation rule from labeled source domains. Candidate relation adjustments are retained for learning this rule when they reduce classification loss on disjoint source validation trials. Consequently, reducing the number of available target trials does not require learning a new relation basis, graph encoder, or context-to-coefficient mapping. The target data are used to infer how a previously source-trained rule should be applied to the current recording.

Second, target context is constructed by aggregating trial-level graph representations using estimated class probabilities. Rather than independently estimating a large collection of target-specific edge parameters, TRACE predicts an eight-dimensional relation-coefficient vector. The coefficients are further constrained to the convex hull of source-domain predictions and the reference coefficient vector. This restricts extrapolation when target context is estimated from fewer trials.

Third, the source-trained graph encoder and within-relation edge proportions remain fixed. Target coefficients modify the relative total weights of the shared spatial–spectral relation types. TRACE then projects the difference between adjusted and reference graph representations into a residual that complements the base EEG representation. Target adaptation therefore uses a constrained relation-adjustment pathway while retaining the source-trained relational structure and the base feature pathway.

These mechanisms provide a structural explanation consistent with the limited-budget results. The experiment evaluates their combined behavior rather than attributing the measured performance difference to any single component.

We also agree that streaming targets are a valuable direction for further investigation. TRACE's trial-level context aggregation suggests a possible route toward incremental context and prototype estimation, while drift-aware coefficient updates could accommodate changes over time. A streaming evaluation would require chronological adaptation and prediction, ensuring that each newly arriving trial is scored before its information is incorporated into subsequent updates. The present experiment establishes the behavior of offline adaptation with smaller unlabeled target batches; it does not assume such a streaming procedure has already been evaluated.

AuthorsOpenReview ↗

## Response to Reviewer uESJ — Part 2/3

W2 and Q2. Statistical reliability and variability across target recordings and random seeds.

The accuracy SD reported in Table 1 describes variation across target recordings. Specifically, for each of the original three seeds, we compute the population SD of recording-level accuracy and then average these SDs across seeds (Section 4.1; Appendix D.1). It therefore captures the heterogeneity of target recordings.

To examine the latter, we conducted a supplementary five-seed evaluation across all eight transfer directions, using five seeds (seeds 20261110–20261114). Within each direction, the target-recording evaluation set is held fixed across seeds. Each seed independently performs source-side learning and target adaptation, using the same per-recording metric definitions and equal-weight recording aggregation as the main evaluation.

For a given seed, we first average each metric across the evaluated target recordings. We then calculate the mean and sample SD of the five seed-wise averages.

Transfer directionAccuracy (%)Macro-F1 (%)BAcc (%)
SEED → SEED-IV54.2551.7355.59
SEED-IV → SEED60.3660.5060.33
SEED → SEED-V60.1362.3960.58
SEED-V → SEED60.2159.6856.78
SEED-IV → SEED-V64.4959.4664.61
SEED-V → SEED-IV64.0057.4656.47
SEED³ → MPED³44.3543.6143.56
MPED³ → SEED³52.6749.4150.57

Across the eight directions, these observations support the consistency of TRACE's results across independently seeded runs on the evaluated target sets. They also clarify why the larger recording-level SDs in Table 1 should not be interpreted as equivalent to run-to-run training instability. Recording heterogeneity and random-seed variability are distinct properties, and the two analyses report them separately.

Paired statistical significance against baselines. This is a separate question from run-to-run variability. A paired comparison requires TRACE and the strongest baseline in each transfer direction to be evaluated on matched target recordings, with the dependence between sessions of the same participant appropriately handled. The supplementary five-seed results address the reproducibility of TRACE's performance estimates.

AuthorsOpenReview ↗

## Response to Reviewer uESJ — Part 3/3

W3. Motivation and interpretation of Figure 1.

Figure 1 uses controlled diagnostics on SEED → SEED-IV to answer two questions motivating TRACE.

First, which spatial–spectral relations should be attenuated? Figures 1(a–b) compare four strategies within the same eligible regional groups: Target selects relations with beneficial attenuation effects on target discovery trials; Source uses source-domain effects with matched selection counts; Random matches these counts by random selection; and Uniform distributes the same total attenuation across candidate relations. Figure 1(a) uses a source-only predictor, whereas (b) uses a target-aligned predictor. Predictor parameters and selected relations remain fixed throughout the attenuation sweep.

Selected relations and predictor parameters remain fixed across the attenuation sweep; selections are made on discovery trials and evaluated on independent confirmation trials.

In both settings, Target yields larger accuracy improvements than matched Random or Uniform controls. Thus, the result depends on which frequency-specific relations are adjusted, not merely on reducing edge weights or selecting more relations.

Second, do useful relation selections transfer across target recordings? Figure 1(c) compares centered relation-effect profiles from disjoint trials of the same target with those from different targets. Their agreement is higher within the same recording, indicating that relation effects are more repeatable across trials from a given target than transferable across recordings. Figure 1(d) evaluates Current (current-target selections), Other (selections from other targets), Pooled (aggregate effects), and Random on held-out trials. Its metric is NLL reduction relative to the unmodified predictor, so larger positive values indicate greater improvement. Current produces a larger reduction than Other. This provides a prediction-level check of the effect-agreement observation in panel (c): selecting a relation because it helps another recording need not give the same held-out benefit on the current recording.

Together, these findings motivate both relation-specific adjustment and target conditioning. Target labels are used for these diagnostics, whereas TRACE infers coefficients from unlabeled target context using a rule learned on labeled sources. We clarified the Introduction and Figure 1 caption; the full protocol appears in Appendix E.1.

Q3. Contribution of graph-based relation modeling beyond target-conditioned adjustment.

Table 2 original results contains two complementary controls. w/o Graph Relation Adjustment (last row) keeps the inferred target-specific coefficients but injects them directly as a feature residual instead of changing spatial–spectral graph weights. w/o Target Conditioning (second row) preserves graph-based adjustment but replaces target-specific coefficients with one shared source-optimized vector. In the first control, the coefficient vector is still inferred from each target, so the difference is not explained simply by having access to target context. In the second, graph-relation weighting is still active, so its performance gap isolates the benefit of target-dependent selection within the tested adaptation pipeline. Thus, the former tests how coefficients enter the model; the latter tests whether their selection depends on the current recording.

Transfer directionFull TRACEw/o Graph Adjustmentw/o Target Conditioning
SEED → SEED-IV54.8250.6050.10
SEED-IV → SEED60.8953.2754.59
SEED-IV → SEED-V64.2655.0653.67
SEED-V → SEED-IV63.3257.9256.27

All values are accuracy (%). Full TRACE gains 4.22, 7.62, 9.20, and 5.40 points, respectively, over w/o Graph Relation Adjustment, and 4.72, 6.30, 10.59, and 7.05 points over w/o Target Conditioning. The former comparison supports explicit graph-mediated adjustment rather than direct feature conditioning. The latter supports recording-specific coefficient inference rather than imposing one source-derived adjustment on every target. The gains occur in all four tested directions, not just in one transfer setting.

Table 2 also includes w/o Relation Adaptation, retaining reference relation weights, and w/o Relational Residual, substituting the reference graph representation for the adjusted-minus-reference difference. All variants use the same adaptation procedure and budget but are adapted separately, so results compare complete adaptation pathways rather than frozen-network interventions. Together, they support distinct, complementary benefits from target-specific coefficients and their graph-mediated application. We clarified this interpretation in Section 4.3.

Reviewer_W2dZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of cross-dataset EEG emotion recognition. It starts from two claims: different target domains require different 'graph relation adjustments', and these adjustments can be predicted from unlabeled target data. On the source side, the authors build pairs of (domain statistics, validated adjustment coefficients) and fit a linear mapping on them. At test time, the mapping predicts an adjustment for the target domain. This adjustment is constrained by a convex-hull rescaling and injected into the downstream model through a residual branch. On eight cross-dataset directions over SEED, SEED-IV, SEED-V and MPED, TRACE outperforms all compared baselines (Table 1).

Critical strengths

S1. Relevant and practical problem. Distribution shift across subjects, sessions and datasets is a central obstacle for EEG-based emotion recognition. The idea of adapting to each target domain individually, rather than learning a single shared alignment, is well motivated.

Critical weaknesses

W1. Large, undisclosed gaps between the re-run baselines and the values reported in the original papers. For HEDN, the original paper reports 69.69 / 73.63 on SEED→SEED-IV / SEED-IV→SEED (its Table IV), while TRACE's Table 1 gives 51.68 / 54.01 1. We suspect this may be partly due to a difference in model-selection protocol. For example, the official HEDN code reports both best-epoch and final-iteration accuracy 2, whereas TRACE uses the final iteration. However, the paper does not discuss this. More importantly, a comparable gap also appears where the setting should be directly comparable. For E2STN on MPED3→SEED3, TRACE states that it follows E2STN's label mapping (Appendix D.1). Yet the original paper reports 73.51 (its Table 1), compared with 48.79 in TRACE's Table 1 [3] (https://arxiv.org/html/2403.16540). This 24.7 pp gap needs an explicit explanation.

W2. Positioning with respect to prior work on conditioning models on unlabeled target data. Adapting or conditioning a predictor based on unlabeled target data has been studied before such as Adaptive Risk Minimization, which the paper cites 4. Meta-learning has also been applied to cross-subject EEG emotion recognition 5. The paper would benefit from a clearer discussion of what is specific to its design compared with these approaches, and why this particular design suits the task. In addition, the method relies on the premise that the required adjustment can be predicted from unlabeled target data. The paper would be stronger with direct evidence for this point.

W3. Inconsistency between the substitution experiment and the motivating analysis. In Figure 1(d), applying another target's adjustment ("Other") still reduces the loss compared with no adjustment. In Table 3, however, "Other" yields lower average accuracy than "Zero": 54.23 vs. 56.81 on IV→V, and 55.59 vs. 57.81 on V→IV. The paper does not discuss this discrepancy.

References

  1. HEDN
  2. HEDN code
  3. E2STN
  4. ARM
  5. MAML for EEG emotion
Questions for authors
  1. For the re-run baselines, especially those with official code such as HEDN, could the authors release the reproduction code, or report both best-epoch and final-iteration accuracy? This would make the effect of the protocol difference much clearer. Could the authors also explain the gap for E2STN on MPED3→SEED3?
  2. The supplementary material contains only inference code, and it does not match the paper. The parameter count of the released model and its computational cost differ from Table 8. Since the code is provided as supporting evidence, it should correspond to the implementation used for the reported results. Please clarify this discrepancy and release the full training code or checkpoints.
  3. In Figure 4, the change in relation weights caused by the target adjustment is on the order of 1e-5 to 1e-4. Removing relation adaptation (Table 2, "w/o Relation Adaptation") reduces accuracy by 2.7–5.7 pp. Could the authors explain how such small weight changes lead to these differences?
  4. In Table 1, the macro-F1 on S→V (66.73) is 7.18 points higher than the accuracy (59.55). In all other directions, F1 does not exceed accuracy. Could the authors confirm that this value is correct?
2 comments
AuthorsOpenReview ↗

Response to Reviewer W2dZ

Thank you for the careful review. We should have made clear that Table 1 reports common-protocol baseline reruns, not published scores. We now clarify this in Section 4.1 and below Table 1. Changes for Reviewer W2dZ are marked in red.

W1/Q1. Why do the HEDN and E2STN results differ from their published scores?

A1. Table 1 adapts separately to each target recording, evaluates all windows at the final iteration without target-label checkpoint selection, and averages recordings and three seeds equally.

Baseline and transferPublished accuracyTable 1 rerun
HEDN, SEED → SEED-IV69.69%51.68%
HEDN, SEED-IV → SEED73.63%54.01%
E2STN, MPED³ → SEED³73.51%48.79%

HEDN's released pipeline selects a checkpoint using target accuracy and evaluates DBSCAN-retained windows; our rerun uses the final checkpoint and all windows. Although its paper describes subject-level evaluation, its script also enumerates 45 subject–session targets.

E2STN's published protocol uses one target subject; our rerun adapts to 45 subject–session recordings and averages them equally. The scores thus follow different protocols (Section 4.1; Appendix D.3, Tables 8–9). Appendix D.4 (Figure 4) shows baseline loss curves over 1,000 iterations. See Q2 for code availability.

Q2. Why does the supplementary code differ from the full model in size?

A2. The review-stage package illustrates target relation-coefficient inference, including graph construction and unlabeled-context aggregation. It omits full source training and downstream adaptation to protect unpublished core work, explaining its different parameter count.

Appendix D.2 (Table 7) specifies the architecture; Appendix D.5 (Table 10) reports 71,762 full-model parameters and 0.343 GMAC/window for graph-representation computation. Sections 3.1–3.4 and Appendices B–D describe the procedures. Upon publication following acceptance, we will release complete code, baseline reproduction scripts, environment specifications, and trained checkpoints (Section 4.1).

W2. What does TRACE learn from unlabeled target context?

A3. Unlike ARM's predictor adaptation and EEG MAML's labeled-shot decoder adaptation, TRACE maps target context to source-validated spatial–spectral relation coefficients. Source adjustments are selected by lower classification loss on disjoint validation trials. Unlabeled target summaries drive coefficient inference within a source-supported convex hull; a relational residual incorporates the graph-representation change. Sections 2 and 3.1–3.3 also distinguish TRACE from ACCNet and VBH-GNN.

In Section 4.4 (Table 3), Own coefficients outperform five same-session donor vectors in both SEED-IV ↔ SEED-V directions. The source-trained mapping and downstream adaptation procedure are held fixed; only the coefficient vector changes.

W3. Why do Figure 1(d) and Table 3 appear different?

A4. They test different interventions. Figure 1(d) diagnoses single-relation attenuations with a frozen predictor via held-out NLL on SEED → SEED-IV. Table 3 substitutes full donor vectors on SEED-IV ↔ SEED-V, reruns adaptation, and measures final accuracy/macro-F1.

A beneficial donor attenuation can therefore coexist with a full donor vector performing below Zero after adaptation. Both analyses favor the current target's own adjustment (Section 4.4; Appendices E.1–E.2).

Q3. How should the scale of relation-weight adjustments be interpreted?

A5. Figure 5 shows qaT(r)−q0(r)q_{\mathbf a_T}(r)-q_{\mathbf0}(r). Each of 325 relation types has reference mass 1/325≈0.003081/325\approx0.00308, so changes of magnitude 10−510^{-5}–10−410^{-4} correspond to 0.325%–3.25% of that reference. With fixed within-type edge shares, Equation 2 gives

=1+325\left[q_{\mathbf a_T}(r)-q_{\mathbf0}(r)\right].$$ These changes enter graph aggregation and the relational residual. Table 2 evaluates this pathway through ablations with target-adaptation reruns (Section 4.3; Appendix F.1). **Q4. Why is macro-F1 higher than accuracy on SEED → SEED-V?** **A6.** Accuracy weights windows equally; macro-F1 weights classwise F1 equally, so class imbalance can yield higher macro-F1. An actual final-iteration target-recording confusion matrix (true classes in rows) is: | True / predicted | Negative | Neutral | Positive | |---|---:|---:|---:| | Negative | 753 | 392 | 0 | | Neutral | 392 | 340 | 0 | | Positive | 25 | 0 | 98 | This yields 59.55% accuracy, 66.73% macro-F1, and 63.96% BAcc, with classwise F1 of 65.05%, 46.45%, and 88.69%. Table 1 calculates these metrics per recording from final-iteration predictions, then averages across 48 recordings and three seeds (Section 4.1; Appendix D.1).

Dear Reviewer W2dZ,

We sincerely appreciate the time and effort you have devoted to reviewing our paper. Your thoughtful and constructive feedback has been very valuable in helping us improve our work.

We have carefully considered each of your comments and provided detailed point-by-point responses, including further clarifications, additional experimental evidence, and corresponding manuscript revisions where appropriate.

As the discussion period is approaching its end, we would be very grateful if you could kindly review our responses and share any further thoughts at your earliest convenience. We would be happy to address any remaining questions and continue the discussion.

If our responses have satisfactorily addressed your concerns, we would also sincerely appreciate your reconsideration of your rating and confidence, as you deem appropriate.

Thank you again for your valuable time and consideration.

Best ,
Authors

Reviewer_tCrsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TRACE for cross-dataset EEG emotion recognition. It learns a mapping from domain context to relation coefficients using source data. These coefficients adjust spatial–spectral relation weights for unlabeled target domains. The method then computes the difference between the adjusted and reference graph representations. This difference supports subsequent adaptation and emotion prediction. Experiments cover eight transfer directions across four datasets and demonstrate strong overall performance.

Critical strengths

1.The method addresses a meaningful research problem. By learning how to adjust relations according to target-domain data, TRACE provides a targeted adaptation approach for cross-dataset EEG emotion recognition.

2.The experiments provide reasonably comprehensive coverage and consistent results. TRACE achieves higher mean accuracy than the compared baselines across all eight transfer directions, with improvements of 2.22–3.94 percentage points over the strongest baseline in each direction. The component ablations in Table 2 and coefficient-substitution experiments in Table 3 also provide substantive evidence supporting target-specific relation adjustment.

Critical weaknesses
  1. Several important experimental details require clarification to improve reproducibility. Figure 1 uses different levels of attenuation, but the quantities being attenuated, the implementation, and the normalization procedure are insufficiently clear. The paper should also clarify how the training pairs are constructed. Each pair consists of a domain context and a corresponding relation adjustment. Specifically, the authors should report the number of candidate splits, the trial composition of each split, the number of retained pairs, and the corresponding retention rates. These details are important for understanding and reproducing the training procedure.

2.The evaluation results and training comparisons require further clarification. The three-class label mapping gives trial ratios of 2:1:1 for SEED-IV and 3:1:1 for SEED-V. Although Macro-F1 is reported, balanced accuracy would provide a useful complementary measure of performance across classes. Additionally, training all methods for 1,000 iterations provides a common protocol but does not ensure that every method has sufficiently converged. The authors should therefore explain the choice of this training budget.

Questions for authors
  1. How is the attenuation corresponding to (\rho\in{0,0.1,\ldots,1.0}) in Figure 1 implemented? What quantities are attenuated, is renormalization applied afterward, and are all other model parameters and experimental conditions held fixed?

  2. The three-class mapping introduces trial-level class imbalance in SEED-IV and SEED-V. Since accuracy gives greater weight to classes with more samples, would balanced accuracy be a more appropriate complementary metric?

  3. When constructing the training pairs ((\mathbf c,\mathbf a)), how many adjustment/validation splits are attempted per source domain? How many trials from each class are included in each subset, and how many pairs are ultimately retained? Please report these details and the corresponding retention rates to support reproducibility.

  4. Have the main baselines sufficiently converged within the common budget of 1,000 iterations? Could the authors provide training curves or results under different training budgets to assess the sensitivity of the performance rankings to this choice?

2 comments
AuthorsOpenReview ↗

Response to Reviewer tCrs

Response to Reviewer tCrs

We thank the reviewer for the constructive feedback. Additions addressing your comments are highlighted in blue in the revised main text and appendix. We invite reconsideration based on these revisions.

Q1. How is attenuation implemented, normalized, and controlled?

A1. We scale selected relation types' pre-aggregation edge weights without global renormalization, with fixed predictors and experimental conditions.

For an edge ee of relation type rr,

we(ρ)=1325mr(ρ)π0(e∣r),w_e(\rho)=\frac{1}{325}m_r(\rho)\pi_0(e\mid r),

where π0(e∣r)=Ae0/∑e′∈ErAe′0\pi_0(e\mid r)=A_e^0/\sum_{e'\in\mathcal E_r}A_{e'}^0 and mr(ρ)=1−ρm_r(\rho)=1-\rho for selected types (and 11 otherwise). The encoder recomputes its usual incoming-message normalization using these weights.

Relations are selected at ρ=0.5\rho=0.5 on discovery trials and evaluated over ρ∈0,0.1,…,1\rho\in\\{0,0.1,\ldots,1\\} on independent confirmation trials. Selected types, trial splits, and predictor parameters remain fixed; source-only and target-aligned predictors are separately trained and frozen. Eligible regional groups have positive attenuation-induced NLL changes on both source and target discovery data. Target selects types with negative discovery effects; Source retains the same groups and counts but ranks types by ascending source effects. Random averages 50 count-matched selections; Uniform attenuates each of mm candidate types by ρk/m\rho k/m when kk are selected. Target labels are used only in this diagnostic, not in TRACE inference.

We have clarified the intervention and controls in the Figure 1 caption and added the full protocol to Appendix E.1 (Eq. 16).

Q2. Does class imbalance affect the conclusions?

A2. The added balanced accuracy (BAcc) results show that TRACE ranks first across all eight transfer directions when the mapped classes receive equal weight.

The mapped trial ratios are 2:1:1 (SEED-IV) and 3:1:1 (SEED-V). BAcc averages three class recalls from final-iteration window predictions within each target recording and seed, followed by equal averaging across recordings and three seeds. TRACE exceeds the strongest baseline in each direction by 1.37–8.69 percentage points, including 55.70% vs. 50.98% on SEED → SEED-IV and 63.96% vs. 55.27% on SEED → SEED-V.

We have added BAcc for all 12 baselines and TRACE to main-text Table 1, and the calculation in Section 4.1 and Appendix D.1. The perturbation analysis separately defines trial-balanced BAcc.

Q3. How are source training pairs constructed and retained?

A3. We now specify the split composition and acceptance criteria and report candidate and retained-pair counts for every direction and seed.

In the final source fit, each seed attempts 16 candidate splits per source domain. Each split uses two adjustment trials and one disjoint validation trial per mapped class (six and three trials in total). A pair (c,a)(\mathbf c,\mathbf a) is retained only if its adjustment strictly reduces loss on both subsets; c\mathbf c is computed from adjustment trials using estimated class probabilities and fixed source prototypes.

We have added the construction procedure to Section 3.4 and Appendix B.2 and the following counts to Appendix Table 6:

DirectionDomainsCand./seedS1S2S3Retention
SEED → SEED-IV4572057959658881.62%
SEED-IV → SEED4572057458556980.00%
SEED → SEED-V4572054755956677.41%
SEED-V → SEED4876859862061379.47%
SEED-IV → SEED-V4572060361459283.75%
SEED-V → SEED-IV4876860562261079.73%
SEED³ → MPED³4572052053351172.41%
MPED³ → SEED³3048036637936076.74%

Retention = (S1 + S2 + S3) / (3 × candidates per seed), ranging from 72.41% to 83.75%.

Q4. Is the 1,000-iteration budget supported by convergence evidence?

A4. We have added training-loss trajectories for all 12 baselines under the common 1,000-iteration budget.

The budget matches HEDN's public configuration (max_iter=1000). All methods are evaluated at the final iteration, avoiding target-label-based checkpoint selection.

We have clarified the budget in Section 4.1 and added the curves and aggregation protocol to Appendix D.4 (Figure 4). They cover SEED → SEED-IV, three seeds, and all 45 recordings for target-adaptive methods. All 12 baselines show substantial loss reduction, with later segments generally changing more slowly; training-phase changes and differing loss scales are indicated. These trajectories and HEDN's public setting provide empirical support for the budget used in Table 1.

Dear Reviewer tCrs,

We sincerely appreciate the time and effort you have devoted to reviewing our paper. Your thoughtful and constructive feedback has been very valuable in helping us improve our work.

We have carefully considered each of your comments and provided detailed point-by-point responses, including further clarifications, additional experimental evidence, and corresponding manuscript revisions where appropriate.

As the discussion period is approaching its end, we would be very grateful if you could kindly review our responses and share any further thoughts at your earliest convenience. We would be happy to address any remaining questions and continue the discussion.

If our responses have satisfactorily addressed your concerns, we would also sincerely appreciate your reconsideration of your rating and confidence, as you deem appropriate.

Thank you again for your valuable time and consideration.

Best,
Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.