acceptodds
Under review as a conference paper at ICLR 2027

Multi-Level Transport Alignment for Time-Series Domain Generalization

Abstract

Time-series domain generalization is difficult because class semantics are often entangled with domain-specific signal variation: the same class may appear differently across subjects, sensors, datasets, and recording conditions. Existing pairwise or contrastive alignment methods typically define relations locally, either by matching selected sample pairs or by normalizing relations separately for each sample, which can miss the global structure induced by class semantics and class-conditional domain variation. We propose Multi-Level Transport Alignment (**Multa**), a balanced optimal transport objective for time-series domain generalization. **Multa** learns a globally organized transport structure in the unit representation space, encouraging class-discriminative representations while suppressing spurious source-domain shortcuts. We further introduce an adaptive alignment-gap reweighting mechanism that focuses optimization on samples whose learned alignment structure deviates most from the target relation. We show that globally balanced optimal transport yields a tighter bound on class-conditional source variation than row-wise local matching for fixed representations, and derive a generalization bound linking Multa’s transport objective to unseen-domain risk. Across five leave-one-domain-out cross-dataset sleep staging tasks and 25 cross-person human activity recognition tasks, **Multa** consistently improves over existing baselines under both dataset-level and subject-level domain shifts.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
domain generalization, optimal transport

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_WtfxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MULTA (Mutli-Level Transport Alignment) for time-series domain generalization. MULTA is a balanced optimal transport objective and learns a globally organized transport structure in the unit representation space. The globally balanced optimal transport yields a tighter bound on class-conditional source variation than row-wise local matching.

Critical strengths
  • The abstract and Introduction section motivate the challenge very well and clarify the contributions of the paper. The paper is easy to understand.
  • The main methodological strength is the mathematically transparent analysis connecting the objective to representation geometry.
Critical weaknesses

Weaknesses:

  1. A visualization of the proposed theoretical alignment would help to understand Section 4 better.
  2. The evaluation contains a comparison to several state-of-the-art techniques. Why are many domain adaptation techniques not included? Most of the baseline methods are 2 years old. Methods such as DAN, CDAN, HoMM, CoDATS, AdvSKM, CPDA, Bures alignment techniques, and also Slices Wasserstein DA are not included. Furthermore, there exist recent architecture-specific time-series DA methods, such as CauDiTS, TransPL, LCA, VLH-DA, DWLR, TA4LS, CADT, TF-DAN, FADA-SPL, CATS, and Rainforest. Furthermore, I think the contribution in the related work section is insufficiently distinguished from the closest prior transport-based and contrastive learning methods.
  3. While the results of MULTA perform significantly better than comparison techniques on the P2018, DSADS, USC, HAR, EMG, and PAMAP datasets, the improvement are marginal for other datasets, including ISRUC (no improvement), Sleep-EDFx, SHH, HMC, etc.
  4. Regarding the ablation study: Component ablations and sensitivity analyses are restricted to EMG and PAMAP2, using three seeds. There are no corresponding ablations on sleep staging, despite its different signal characteristics and dataset-level shifts.
  5. The limitations section state that “MULTA introduces additional computational overhead due to optimal transport estimation”. How large is the overhead compared to state-of-the-art techniques? A comparison of training/inference times and RAM usage would be necessary. The batchwise transport matrices require quadratic storage, yet the paper provides no controlled runtime or peak-memory comparison against ERM and SupCon.

Minor:

  • In Figure c (e to h), the axis labels and unit are missing.

Reproduciblity:

  • I had a look on the source code and appendix and believe that the paper is reproducible. Datasets are publicly available.
  • Appendix I uses pairwise, two-sided paired t-tests over only five seeds, without a reported multiple-comparison correction. With approximately 20 methods, this produces many comparisons per task -> however, the five training runs provide limited support for assessing distributional assumptions.
Questions for authors

See the questions in the weaknesses

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

This paper studies zero-shot domain generalization for time series, where class semantics are entangled with domain-specific signal variation, and existing pairwise or contrastive methods define relations only locally, missing the global structure induced by class-conditional domain variation. The authors propose MULTA, a balanced optimal transport objective that learns a globally organized transport structure, suppresses spurious source-domain shortcuts, and adaptively reweights samples deviating most from the target relation, with theoretical guarantees including a tighter class-conditional source variation bound than row-wise local matching and a generalization bound linked to unseen-domain risk. Across 5 cross-dataset sleep staging and 25 cross-person HAR tasks, MULTA consistently improves over baselines under both dataset-level and subject-level domain shifts.

Critical strengths
  1. The paper pinpoint a real limitation of existing time-series DG methods—pairwise alignment without control over the global cross-domain organization of samples—and position their contribution clearly against it.

  2. MULTA casts class-aware alignment as distribution-level optimal transport, enabling explicit global control of cross-domain relational structure; the transport alignment-gap weighted objective is accompanied by theoretical analysis linking the matching loss to class-conditional source variation and an unseen-domain risk bound.

  3. Experiments span SleepDG cross-dataset sleep staging (dataset-level shifts) and HAROOD cross-person HAR (subject-level shifts), covering multiple domains, tasks, and seeds.

Critical weaknesses

On the sleep staging datasets, MULTA outperforms the strongest baseline by only 1.17 points on average (72.22 vs 71.05) and even falls behind ERM++ on ISRUC; the 3.01-point gain on HAROOD is concentrated on the high-variance WESAD dataset, with substantial fluctuation across seeds. The claim of consistently improves is stronger than what the data actually support.

The paper does not construct a minimal variant (e.g., “SupCon + domain-aware positive-sample weighting, without Sinkhorn”) to disentangle the contributions of “global transport” and “domain-aware positive-sample weighting.”

Component ablations cover only 2 of the 6 HAROOD datasets (with only 3 seeds); the method has roughly 8-9 hyperparameters, yet sensitivity analysis is performed only for \lambda_tr and \lambda_h; setting \lambda_tr=2 in fact degrades performance, suggesting a narrow usable hyperparameter range.

Questions for authors

Please answer the questions in the weakness section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NBpn

We thank the reviewer for this constructive feedback which greatly improved the clarity and quality of our original manuscript. We have expanded the baseline methods as well as the number of tasks and the random seed number for the ablation and sensitivity analyses with the manuscript being updated as well.

W1. Magnitude and consistency of improvements

We agree that “consistently improves” should distinguish aggregate performance from improvements on every individual target. MULTA achieves the highest average balanced accuracy on both benchmarks, although the magnitude of improvement varies across datasets. We have revised the wording in the abstract, introduction, experiments and etc. to make this distinction explicit.

To strengthen this comparison, we have added four recent methods with public code that cover complementary approaches: COGS (AAAI 2026) for causal representation learning [1], DCCL (KDD 2025) for domain-connecting contrastive learning [2], CorSW (KDD 2026) for sliced-Wasserstein alignment of correlation distributions [3], and PhASER (TMLR 2025) for phase-aware learning in nonstationary time series [4].

Sleep cross-dataset DG

MethodP2018ISRUCSleep-EDFxSHHSHMCAvg.BH sig. rank ↓
DCCL73.05 ± 1.0470.45 ± 1.3568.06 ± 1.0766.00 ± 0.6370.99 ± 0.7969.71 ± 0.403.60
CorSW72.13 ± 1.3670.89 ± 1.3168.87 ± 1.2766.88 ± 2.2469.31 ± 2.0769.62 ± 0.972.60
PhASER66.04 ± 0.8362.59 ± 1.4362.72 ± 0.3551.68 ± 2.6661.10 ± 0.5160.82 ± 0.2519.60
COGS68.86 ± 3.9266.43 ± 3.3665.53 ± 3.6763.03 ± 2.5565.32 ± 2.0665.83 ± 2.126.60
MULTA75.00 ± 0.7572.56 ± 0.6372.16 ± 0.5568.63 ± 0.9672.77 ± 0.8772.22 ± 0.331.00


HAROOD cross-person DG

MethodDSADSUSCUCI-HAREMGWESADPAMAPTask Avg.BH sig. rank ↓
DCCL83.89 ± 0.8670.02 ± 0.5592.64 ± 0.1857.51 ± 3.4626.44 ± 2.4266.14 ± 0.5867.17 ± 0.327.94
CorSW81.38 ± 1.2169.42 ± 0.2592.56 ± 0.1277.90 ± 0.4343.60 ± 1.8665.46 ± 1.2672.55 ± 0.283.84
PhASER85.34 ± 1.7167.92 ± 1.3993.75 ± 0.2275.52 ± 1.0748.27 ± 0.4970.95 ± 1.2274.43 ± 0.323.76
COGS82.20 ± 0.8174.26 ± 2.2192.75 ± 0.4657.82 ± 9.2628.89 ± 4.2067.87 ± 1.4468.31 ± 2.375.61
MULTA88.35 ± 0.8474.62 ± 0.4694.48 ± 0.2780.46 ± 0.2150.82 ± 2.6871.56 ± 1.0577.43 ± 0.261.00

Both tables report balanced accuracy in the form of mean ± SD across the same five seeds as the main paper. We also apply the Benjamini–Hochberg (BH) correction[5] within each target task to control false positives from multiple testing while retaining our original significance rank definition.

Across both benchmarks, MULTA still achieves the highest average balanced accuracy and maintains the best BH-corrected significance rank of 1. These comparisons support its effectiveness alongside recent causal, contrastive, transport-based, and time-series-specific DG approaches.

Theorem 1 also clarifies the limits of our claim: MULTA addresses variation across source domains, while the bound separately includes the unseen domain’s discrepancy from its closest source distribution mixtures. Source-only training cannot measure or guarantee a reduction in this term, so the analysis does not imply uniform gains across unseen targets.

Reviewer_P8qkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Multi-Level Transport Alignment (MULTA), an optimal transport-based approach for time-series domain generalization. The method aims to improve generalization to unseen domains by aligning representations across classes and source domains while preserving class-discriminative information. It also introduces an adaptive alignment-gap reweighting mechanism to prioritize samples with larger alignment discrepancies.

The authors provide theoretical analysis supporting the proposed alignment objective and evaluate the method on sleep staging and human activity recognition tasks. Experiments across 30 domain generalization tasks, including comparisons with existing methods, ablation studies, and sensitivity analyses, show improvements in balanced accuracy over the evaluated baselines.

Critical strengths

The paper provides an extensive experimental evaluation across 30 domain generalization tasks, covering both sleep staging and human activity recognition. The comparison with multiple baselines, together with ablation studies, sensitivity analyses, and results across different random seeds, strengthens the empirical evidence supporting the effectiveness and robustness of the proposed approach.

The authors provide mathematical propositions and a generalization bound connecting the proposed transport alignment objective to class-conditional source variation and unseen-domain risk. This theoretical analysis is a valuable contribution, as it provides a mathematical justification for the proposed approach beyond empirical performance improvements.

Critical weaknesses

Line 074: It would be better to introduce the full name of the method or concept before using its abbreviation, as was done in Line 090. This would improve readability and maintain consistency throughout the paper.

Figure 1: The figure contains too much information, making it difficult to interpret. Additionally, some abbreviations and symbols are not clearly defined, which further reduces its readability. The authors should consider simplifying the figure, splitting it into smaller figures, or providing a more detailed caption explaining each component.

Lines 195–196: The meaning of the subscripts id, sd, and cd is not immediately clear. Although these terms can be inferred from the preceding description, they should be explicitly defined when first introduced to make the mathematical formulation easier to follow.

Line 224: The authors state that the coefficients β⁺ and β⁻ are non-negative and that their sum must be strictly positive. However, the motivation for this constraint is not explicitly explained. The authors should clarify its role in the alignment-gap formulation and explain why at least one of the coefficients must be positive.

Questions for authors

The authors should provide an experimental comparison between the full MULTA model, its version without alignment (w/o Align), and competitive baselines such as SupCon and InfoNCE. This comparison should report training time, peak GPU memory consumption, and predictive performance under the same hardware and experimental conditions. Such an analysis would help determine whether the performance improvements achieved by MULTA justify the additional computational cost introduced by optimal transport.

2 comments
AuthorsOpenReview ↗

Response to Reviewer P8qk

We thank the reviewer for the very helpful feedback which greatly improved the readability and quality of our original manuscript. We have improved the revised manuscript following your advice and also measured the training and inference costs. In general, MULTA improves over SupCon by 1.35 points on SleepDG and 3.01 points on HAROOD on average, while adding only 1.41% training time over SupCon and no measurable inference cost.

W1. Abbreviations at first use

We thank the reviewer for pointing this out. We have updated Line 66 to include the full names before their abbreviations.

W2. Figure 1 readability

Figure 1 compares the target plan, the row-normalized plan, and MULTA’s learned coupling, while the bottom panels show the corresponding distributions for individual anchors. To improve readability, we have expanded the caption to define all sleep-stage abbreviations (W, N1–N3, REM), domain symbols (dkd^{k}), matrix and line colors, and shaded target-support regions as well as improving the wording and expressions to make the figure self-contained and easier to follow.

W3. Definitions of id, sd, and cd

id\mathrm{id} denotes identity pairs, where a sample is paired with itself. sd\mathrm{sd} denotes distinct samples with the same class and source domain. cd\mathrm{cd} denotes samples with the same class but different source domains. Their coefficients determine the relative scores used to construct the target plan before balancing. We now define all three subscripts at first use in the opening paragraph of Section 3.3.

W4. Constraints on β+\beta^{+} and β−\beta^{-}

The two gaps measure different alignment errors. The first measures insufficient mass on same-class cross-domain candidates, and the second measures mass assigned to different-class candidates. The nonnegative coefficients make the combined score a weighted average of these two gaps so that increasing either error cannot decrease the score. Their sum must be positive so that the denominator is nonzero and this average is well-defined. Either coefficient can be zero if only the other gap is used. We have added this explanation in Section 3.4 in color blue text.

Q1. Computational cost and predictive performance

We evaluated with P2018 held out, using the same five random seeds, a batch size of 640, FP32 precision, and the same 45-GB NVIDIA B200 MIG partition. We report predictive performance as balanced accuracy (BA) on P2018. BA and host RAM (Slurm MaxRSS ) are reported as mean ± SD across the five seeds.

MethodP2018 BA (%) ↑Training time (ms/update)Peak GPU memory (GiB)Host RAM reference (GiB)Inference time (ms/batch)
ERM (w/o Align)70.47 ± 1.76336.19 ± 0.2330.45975.12 ± 1.58113.244 ± 0.187
SupCon / InfoNCE73.21 ± 0.41338.79 ± 0.2830.46064.64 ± 0.36113.273 ± 0.274
CORAL69.82 ± 1.38341.52 ± 1.0130.46164.47 ± 0.20113.262 ± 0.277
MULTA75.00 ± 0.75343.56 ± 0.9630.46065.27 ± 0.22113.243 ± 0.193
PCL73.24 ± 1.16344.97 ± 0.2930.778129.10 ± 3.70113.244 ± 0.192
MMD71.18 ± 0.93350.04 ± 1.3730.46165.32 ± 0.95113.266 ± 0.246
LAG71.30 ± 1.73391.92 ± 0.1729.26994.55 ± 1.21133.644 ± 0.290
RSC72.37 ± 1.53484.89 ± 32.1631.13455.03 ± 0.35113.286 ± 0.272
SleepDG72.96 ± 0.62636.50 ± 0.5933.96194.60 ± 0.14113.244 ± 0.170

Balanced accuracy and training cost. On P2018, MULTA improves mean BA over SupCon by 1.79 percentage points at an additional 4.77 ms per update (1.41%), with the same measured peak GPU memory of 30.4606 GiB. Compared with ERM (w/o Align), MULTA gains 4.53 percentage points at an additional 7.37 ms per update (2.19%).

For batch size BB and feature dimension dd, we compute the affinity matrix once per update in O(B2d)O(B^2 d) time. Each Sinkhorn iteration rescales this matrix, adding O(IB2)O(I B^2) time for II iterations.Storing the relation and transport matrices requires O(B2)O(B^2) memory which is the same for contrastive methods.

Inference. MULTA uses only the encoder and classifier at test time. MULTA and ERM both use 4.6196 GiB of forward-pass GPU memory and take approximately 113.24 ms per batch.

I thank the authors for their detailed responses and for conducting the additional computational cost experiments.

The clarifications regarding the mathematical formulation and the revisions addressing presentation concerns adequately resolve my initial comments. In particular, the additional comparison of predictive performance, training time, and GPU memory consumption provides useful evidence that MULTA achieves improvements over SupCon with relatively modest computational overhead.

Although the computational analysis is limited to one held-out domain, I consider the response satisfactory and believe that my main concerns have been adequately addressed. I therefore maintain my original Weak Accept recommendation.

Comments on the paper

2 comments
Area_Chair_sKT5OpenReview ↗

Please engage with the authors rebuttal by Nov 18

Dear Reviewers,

The authors have posted their rebuttal. Please engage with it by November 18, 2026, noting which of your concerns are resolved or which remain. Even brief engagement helps greatly.

Best regards,
Area Chair

AuthorsOpenReview ↗

General Response

Dear Reviewers, AC, SAC, and PC,

Thank you for your time and effort in reviewing and handling our paper. We are pleased that the reviewers highlighted the paper’s clear motivation and the theoretical link between the transport objective and representation geometry (Reviewer Wtfx). On method and novelty, Reviewer NBpn emphasized how MULTA goes beyond pairwise alignment by using balanced optimal transport to jointly align same-class samples across source domains. Reviewer P8qk valued the theoretical bound linking class-conditional source variation to unseen-domain risk and the empirical evaluation across 30 held-out tasks in sleep staging and human activity recognition. We summarize the discussion and then highlight the new experiments and modifications to the manuscript.

Summary of discussion:

During the rebuttal period, all three reviewers replied that their concerns were addressed and maintained their scores at 3 (Weak Accept).

  • Reviewer P8qk confirmed that their "main concerns have been adequately addressed".

  • Reviewer NBpn wrote that the SupCon + reweighting ablation “convincingly disentangles the contributions of global Sinkhorn matching and the alignment-gap reweighting.” They also described our expanded sensitivity analysis across six datasets with five seeds as “substantially stronger than that in the original submission.” In response to their follow-up questions, we explained how we selected the number of Sinkhorn iterations, how we set the target-plan weights, and how we selected the transport weight using source-domain validation. The reviewer concluded, “I have no further questions and will maintain my current rating.”

  • Reviewer Wtfx wrote that the new baselines, expanded ablations and hyperparameter analysis, and runtime and memory comparisons “address my main empirical concerns.”

Overview of results and changes to the manuscript:

  • Additional baselines (NBpn, Wtfx). We added four recent domain generalization baselines (DCCL, CorSW, PhASER, and COGS). MULTA retains the highest average balanced accuracy and all of the new baselines underperform relative to our previous baselines.

  • Expanded ablations, sensitivity and hyperparameter analysis (NBpn, Wtfx). We expanded the ablations, added additional tasks to our sensitivity analysis, and increased from three to five seeds per task. We also swept the number of Sinkhorn iterations and found that our original settings remained optimal.

  • Training and inference costs (P8qk, Wtfx).: We measured training and inference costs. MULTA gains 1.79 balanced accuracy points over SupCon with 1.41% more time per training update, the same peak GPU memory, and no measurable inference overhead.

  • BH correction for multiple statistical significance tests (Wtfx): We applied the Benjamini–Hochberg (BH) correction to limit false positives when comparing many methods on each task. After correction, MULTA retained the top average rank over all methods.

  • Manuscript revision (P8qk, Wtfx): We revised the text and added more figures as the reviewer requested: we clarified the plans and axes in Figure 1. We added a new Figure 2 in the Appendix that illustrates two terms in the unseen-domain risk bound, the same-class source variation that alignment can reduce and the distance from the target class distribution to its closest source mixture.

Overall, we are happy that all the reviewers agreed upon acceptance. The paper is improved through the reviewer feedback and we have incorporated the completed analyses and clarifications into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.