acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Spatial Alignment in UDA for Wearable Multivariate Time Series: A Diagnostic Analysis and a Simple Remedy

Abstract

Wearable sensing models often fail when deployed on new users, devices, or sensor placements, because the meaning of each sensor channel is inconsistent across these settings. Unsupervised domain adaptation (UDA) for multivariate time series (MTS) aims to close this gap, yet existing methods mainly address temporal distribution shift or align domains using spatial structure learned within each domain separately. Through a diagnostic analysis, we identify a distinct failure mode in wearable MTS, which we term sensor-wise domain shift. Current alignment strategies do not model this cross-domain channel mismatch explicitly and therefore struggle to correct it. We propose (Inter-Domain Sensor Alignment), a backbone-agnostic plug-in module that aligns sensors explicitly across domains. combines (i) inter-domain sensor transport, which learns a source–target sensor correspondence matrix from domain-specific sensor embeddings and maps target channels into the source sensor space, and (ii) channel decorrelation, a regularizer that suppresses redundant or noisy channel couplings and prevents representation collapse. The resulting spatial transportation loss can be interpreted as a sensor-level discrete optimal-transport objective. improves average activity-classification performance over recent baselines on HAR and sEMG benchmarks across most evaluated scenarios.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series, Domain Adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uxrkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to solve sensor-wise domain shift in unsupervised domain adaptation for wearable multivariate time series, where channel semantics change across users or wearing conditions. Specifically, the authors propose Inter-Domain Sensor Alignment (IDSA), which learns a source–target transport matrix to align target channels with the source sensor space, and a channel decorrelation regularizer to preserve distinguishable representations. Experiments on Opportunity, WISDM, NinaPro DB5, and SD-gesture demonstrate improved average classification accuracy and F1-score across four UDA backbones, together with stronger performance preservation under controlled sensor permutations than the evaluated baselines.

Critical strengths
  1. The paper addresses a practically relevant obstacle to reliable wearable sensing.

  2. The design is feasible and easy to follow.

  3. The evaluation includes real public datasets spanning inertial sensing and sEMG, ten main UDA baselines, and additional OT comparisons.

Critical weaknesses
  1. The diagnostic setup allows rotation and scaling, but IDSA’s mixing matrix is much more restrictive.

  2. The OT proof does not show that joint training will find the correct sensor mapping.

  3. The paper does not quantitatively test whether the sensor mapping is correct.

  4. The validation setup is unclear, especially whether tuning uses target labels.

  5. IDSA uses target data, while the foundation-model baselines do not.

Questions for authors

see above

6 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1 (Restricted alignment capacity).

We thank the reviewer for the constructive feedback. We address each point below.

We acknowledge that some of our original statements regarding the transformations that IDSA can handle were broad. To clarify the scope of our method, we conduct additional controlled experiments using synthetically generated datasets, where specific sensor-wise transformations can be independently introduced.

Experimental Results

ScenarioPlainIDSAA=IA = I
Aligned (no shift)90.392.992.9
P: channel permutation37.592.943.1
PF: permutation + sign flips17.36.412.7
F: sign flips only17.211.77.9
D: dead channels87.490.191.0

As discussed in the motivation section, IDSA is primarily designed to address sensor-wise variations involving channel permutation and changes in sensor correspondence. The results support this design objective: under channel permutation (P), IDSA substantially improves accuracy from 37.5% to 92.9%, approaching the performance of the aligned setting (9%). Furthermore, the results under the dead-channel setting (D) suggest that IDSA can maintain or recover useful predictive information even when some sensor channels become unavailable.

However, the results also reveal an important limitation. Under sign-flip transformations (F and PF), where the target sensor signals are inverted (i.e., −x-x instead of xx), IDSA does not provide the expected improvements and performs worse than the original models. We hypothesize that this limitation is related to the formulation of the inter-domain sensor transformation, which represents target sensor information through combinations of existing sensor signals based on source-target relations. Although this formulation can capture changes in sensor correspondence, it may not adequately accommodate transformations involving sign inversion.

These findings help distinguish the transformations that IDSA can effectively address from those that remain challenging for its current formulation. Specifically, IDSA demonstrates strong performance under channel permutation and promising behavior under dead-channel conditions, whereas sign-flip transformations remain a limitation. We acknowledge that our original description did not sufficiently distinguish these cases and will revise the manuscript to clarify the scope and assumptions of IDSA accordingly.

Additionally, to directly address the reviewer’s request, we evaluate Deep CORAL with IDSA under the three transformation settings used in the original diagnostic analysis in Fig. 2(a), namely sensor orientation, sensor placement, and channel permutation. The results show improvements over Deep CORAL in most permutation settings and across several sensor placement conditions. Under axis rotations, however, the gains are mixed. IDSA improves upon Deep CORAL at 108°, 144°, 180°, and 216°, but reduces accuracy at the two smallest tested rotation angles 36° and 72°. Its accuracy also varies less across the tested rotation angles than that of either SASA or Deep CORAL. These results demonstrate that the practical benefit of IDSA depends on the transformation family and that its effectiveness under channel permutation does not imply consistent improvements under general orientation changes. We have updated Fig. 2(a) in the main manuscript to include these additional results and revised the accompanying discussion in Section 4.2 accordingly.

AuthorsOpenReview ↗

Response to Critical Weakness 2 (Uncertain correspondence objective).

Alternative cost analysis

We additionally investigated whether IDSA can utilize alternative cost functions for estimating sensor-wise relationships. Specifically, we considered four alternatives based on different sensor characteristics: eq6, which corresponds to the temporal-profile cost in Eq. (6); logpsd, which measures differences between log-transformed power spectra; pshape, which compares normalized spectral distributions using the Hellinger distance; and combo, which combines spectral information with the mean signal level. For the spectral-based costs, we computed the mean periodogram across samples after removing the temporal mean of each window. We evaluated on the SD-gesture dataset by forcing permutation on the target domain. The accuracy is the mean value of applying ACON and SASA as the baselines.

CostCost HungarianIDSA AccuracyOriginal
logpsd1.0089.646.9
pshape0.5046.446.9
combo0.5652.246.9
eq60.2556.046.9

The results suggest that IDSA's effectiveness depends on the quality of the cost function. In particular, logpsd achieves a Hungarian matching cost of 1 and an accuracy of 89.6%, whereas alternative costs with higher matching costs yield smaller improvements over the Original performance. These findings indicate that IDSA accommodates different cost functions, provided they capture meaningful sensor-wise relationships. Accordingly, we will clarify that Eq. (6) is one possible choice for constructing the sensor-wise cost matrix, rather than a requirement, and that alternative costs based on other sensor characteristics can also be incorporated.

Heterogeneous measurement and phase variation

We further investigated the robustness of IDSA under heterogeneous measurement and phase variation. For heterogeneous measurement, we introduced an additional constant offset to the signals of a specific target domain, considering offset magnitudes of 1, 10, and 100. For phase variation, we shifted the temporal signals in the target domain. We evaluated each scenario both with and without sensor permutation to examine the combined effects of measurement variations and changes in sensor correspondence.

ACON

Permute / modelNo / origNo / IDSAYes / origYes / IDSAHungarian Cost
Offset (100)5.37.26.56.21.0
Offset (10)27.329.011.028.01.0
Offset (1)71.872.046.164.21.0
Shift88.088.657.487.51.0

SASA

Permute / modelNo / origNo / IDSAYes / origYes / IDSAHungarian Cost
Offset (100)6.06.95.55.71.0
Offset (10)25.725.513.717.31.0
Offset (1)73.377.833.573.61.0
Shift91.392.136.791.51.0

IDSA has a limited effect under drastic measurement offsets (e.g., 100), where distribution discrepancies are substantial, but provides notable improvements under moderate offsets combined with sensor permutation. For example, at offsets of 10 and 1, ACON improves from 11.0% to 28.0% and from 46.1% to 64.2%, respectively, while SASA improves from 13.7% to 17.3% and from 33.5% to 73.6%.

Under phase variation combined with sensor permutation, IDSA also substantially recovers performance. Overall, these results suggest that IDSA can mitigate sensor-wise discrepancies arising from changes in sensor correspondence, including those accompanied by moderate measurement offsets or temporal phase variations, although its effectiveness is limited under severe measurement offsets.

Activity Proportions

To investigate the effect of source–target class-distribution differences, we conduct a controlled experiment on SD-gesture by sampling 100 instances from each domain to construct synthetic binary classification settings. We vary the source class proportions as r:(1-r) and reverse them in the target domain to (1-r):r, thereby examining IDSA under different degrees of class-distribution discrepancy. The resulting accuracy (%) is reported below.

SD-gesturerOriginal – no IDSAIDSA
ACON0.586.088.2
ACON0.478.479.6
ACON0.360.661.4
ACON0.231.429.6
SASA0.596.698.6
SASA0.493.895.2
SASA0.379.277.4
SASA0.261.257.8

IDSA achieves comparable performance under moderate distribution differences (r=0.4), but slightly underperforms the baseline under larger differences (r=0.3 for SASA and r=0.2 for ACON). This may be because the mini-batch optimization of Eq. 6 uses averaged representations to estimate sensor-wise dependencies, which can be affected by differing class compositions across source and target batches.

We note that this finding arises from a synthetic setting with explicitly reversed class proportions. In contrast, WISDM exhibits natural class imbalance across subjects (Appendix A.7.1), yet IDSA improves performance in the main evaluation, suggesting that mitigating sensor-wise domain shifts remains beneficial under realistic class-imbalance conditions.

AuthorsOpenReview ↗

Response to Critical Weakness 3 (Incomplete mechanistic validation).

We thank the reviewer for this helpful comment. We acknowledge that the original description of the PCA distributions did not provide sufficient details, and the experimental design in Appendix A.5.2 incorrectly described the transformed target signals as weighted combinations of source signals. In the actual diagnostic procedure, source and target sensor signals are used to estimate cross-domain relations, while the weighted aggregation is performed exclusively over target-domain signals. In the revised manuscript, we have corrected and expanded the explanation to clarify how the cross-domain relations and aggregation are constructed. Specifically, source-domain signals are used to determine the cross-domain relations, whereas the signals being aggregated are exclusively drawn from the target domain. We further clarify that this motivating example illustrates a naive approach that follows a similar underlying principle to our method, in which cross-domain relations are established using source-domain information and subsequently applied to target-domain signals. We believe these additional explanations provide a clearer understanding of both the experimental setup and its connection to our proposed approach.

Additionally, to further examine the effectiveness of our method design, we conduct additional experiments comparing the learnable inter-domain sensor relation matrix Ast\textbf{A}_{st} against two alternative sensor-matching strategies: (1) identity mapping, which replaces Ast\textbf{A}_{st} with an identity matrix I\textbf{I}, assuming direct one-to-one sensor correspondence; and (2) Hungarian assignment, which establishes one-to-one sensor correspondence through dataset-level assignment.

1. Identity mapping (Ast=I\textbf{A}_{st}=\textbf{I}).

Nina18

ModelOriginalChange Identity
IDSA + Deep Coral85.5381.25
IDSA + CLUDA81.8875.76
IDSA + ACON83.7783.57
IDSA + SASA84.3284.22

Opportunity HAR

ModelOriginalChange Identity
IDSA + Deep Coral81.4375.63
IDSA + CLUDA84.6880.25
IDSA + ACON81.9673.99
IDSA + SASA82.9671.16

The results show that replacing the learned Ast\textbf{A}_{st} with an identity matrix generally reduces performance across both Nina5-18 and Opportunity HAR. The performance degradation is particularly noticeable on Opportunity HAR, where accuracy decreases by 4.43–11.80 percentage points across the evaluated backbones. These results suggest that assuming direct sensor correspondence may be insufficient to capture the sensor-wise variations between domains.

2. Hungarian assignment.

Nina18

ModelOriginalChange Assignment
IDSA + Deep Coral85.5375.70
IDSA + CLUDA81.8881.80
IDSA + ACON83.7782.88
IDSA + SASA84.3282.36

Opportunity HAR

ModelOriginalChange Assignment
IDSA + Deep Coral81.4369.62
IDSA + CLUDA84.6874.70
IDSA + ACON81.9679.38
IDSA + SASA82.9676.85

We further compare IDSA with Hungarian assignment, which provides an alternative based on explicit one-to-one direct sensor matching. The results show that the learned Ast\textbf{A}_{st} achieves higher accuracy across all reported comparisons on both datasets. In particular, the performance differences are more pronounced on Opportunity HAR, where Hungarian assignment leads to substantial degradation for both DeepCORAL and CLUDA.

3. Simple Linear Adapter.

OrigIDSAIDSA-MLP
Nina18-deepcoral83.4785.5379.47
Nina18-acon82.0783.7782.05
Nina18-sasa83.0684.3282.41
Nina18-cluda80.6481.8879.97
Oppo-deepcoral78.2781.4375.50
Oppo-acon77.5981.9679.33
Oppo-sasa80.5182.9680.69
Oppo-cluda82.2084.6884.44
SD-deepcoral73.0280.1561.22
SD-acon76.9680.8077.30
SD-sasa75.4577.0174.82
SD-cluda72.4975.6853.68
WISDM-deepcoral62.0175.6373.42
WISDM-acon77.6179.8979.37
WISDM-sasa73.9477.5176.11
WISDM-cluda75.9882.1474.0

From the table above, replacing the Ast\textbf{A}_{st} with an MLP layer consistently degrades performance, suggesting that the proposed IDSA design is effective in mitigating sensor-wise domain shifts.

Taken together, these experiments suggest that learning Ast\textbf{A}_{st} provides advantages over predefined identity correspondence and dataset-level one-to-one sensor assignment. Unlike these alternatives, IDSA learns inter-domain sensor relationships that are not restricted to fixed or one-to-one correspondence. This flexibility appears beneficial for capturing sensor-wise variations across domains and supports the design choice of using a learnable sensor relation matrix in IDSA.

We have incorporated these additional experiments into the revised manuscript to provide further evidence of the contribution of the proposed inter-domain sensor relation learning mechanism.

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

This paper studies sensor-wise domain shift in wearable multivariate time-series UDA, where channel indices may represent different physical signals across domains. It proposes IDSA, a plug-in module combining domain-specific sensor embeddings, a cross-domain channel transport matrix, and transportation and decorrelation losses. Experiments cover four datasets and four underlying UDA methods.

Critical strengths
  1. Consistent improvements across several adaptation objectives. IDSA improves average performance across four underlying UDA methods on both HAR datasets. For example, CLUDA’s accuracy increases from 82.20% to 84.68% on Opportunity and from 75.98% to 82.14% on WISDM. The sEMG evaluation reports results over five shared seeds.
  2. Relevant diagnostics and ablations. The controlled perturbations, component ablations, and transport visualizations provide useful supporting evidence. Tables 3 and 5 show that the proposed losses contribute beyond adding architectural capacity.
Critical weaknesses

The specific contribution of learned cross-domain channel transport is not fully isolated: The existing ablations support the complete method and its losses. However, an identity-transport control retaining the sensor embeddings, GNN, and decorrelation regularization would help establish how much of the improvement specifically requires learning cross-domain channel correspondence.

Questions for authors

Can you provide an identity-transport control while keeping the sensor embeddings, GNN, and decorrelation regularization unchanged?

3 comments

We thank the reviewer for the constructive feedback. We address each point below.

Critical Weakness 1 & Question 1: The specific contribution of learned cross-domain channel transport is not fully isolated: The existing ablations support the complete method and its losses. However, an identity-transport control retaining the sensor embeddings, GNN, and decorrelation regularization would help establish how much of the improvement specifically requires learning cross-domain channel correspondence. & Can you provide an identity-transport control while keeping the sensor embeddings, GNN, and decorrelation regularization unchanged?

We further conduct an intervention experiment to examine the contribution of the learned inter-domain sensor relation matrix Ast\textbf{A}_{st}. Specifically, we replace the learned Ast\textbf{A}_{st} with an identity matrix I\textbf{I}, while keeping the remaining model components unchanged. This allows us to investigate whether learning inter-domain sensor relationships provides benefits beyond assuming direct sensor-to-sensor correspondence.

DatasetModelOriginalChanged Identity
Nina18IDSA + Deep Coral85.5381.25
Nina18IDSA + CLUDA81.8875.76
Opportunity HARIDSA + ACON81.9673.99
Opportunity HARIDSA + SASA82.9671.16

The results show consistent performance degradation when the learned Ast\textbf{A}_{st} is replaced with I\textbf{I} across all evaluated datasets and backbones. For example, on Nina5-18, the accuracy decreases from 85.53% to 81.25% with DeepCORAL and from 81.88% to 75.76% with CLUDA. Similarly, on Opportunity HAR, the accuracy decreases by 7.97 and 11.80 percentage points with ACON and SASA, respectively.

These findings suggest that the learned inter-domain sensor relation matrix contributes meaningfully to IDSA's performance. In particular, the improvements over the identity-based variant indicate that assuming fixed one-to-one sensor correspondence may be insufficient, and that explicitly learning inter-domain sensor relationships can better accommodate sensor-wise variations between domains.

We have incorporated this additional experiment into Section 6.3, entitled "Impact of Inter-domain Sensor Relation Matrix Ast\textbf{A}_{st}," to further investigate the role of the learned sensor relationships in IDSA.

Thank you for adding the identity-transport ablation in Section 6.3. The consistent accuracy drops across the four evaluated configurations when the learned matrix is replaced with the identity matrix address my main concern. My overall recommendation remains 3.

We sincerely thank the reviewer for the follow-up and for acknowledging that the additional ablation addresses the main concern. We appreciate your constructive feedback, which helped us strengthen the experimental validation of IDSA.

Reviewer_Ti8fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors improve unsupervised domain adaptation for wearable signals by addressing changes in channel mapping across domains (e.g., sEMG sensor placement). They propose IDSA, a method that learns a matrix to transform target channels into the source channel space using unlabeled target data. IDSA uses domain-specific sensor embeddings to construct the matrix. A transportation loss encourages mappings between channels with similar signal profiles, while a channel decorrelation loss prevents channel representations from becoming overly similar. These losses are jointly optimized with the existing UDA objective. Experiments on HAR and sEMG datasets show that adding IDSA improves the performance.

Critical strengths
  • The paper addresses a practical challenge in wearable UDA. I like the method which is well motivated by the physical characteristics of wearable sensing, which is valuable for real-world applications.
  • The paper provides a clear connection between its intuition, mathematical formulation, and method design.
  • The comparison with foundation models is interesting. It suggests that adaptation using unlabeled target data can remain useful alongside pre-trained models in practical wearable applications.
Critical weaknesses
  • Importantly, it is unclear whether the learned matrix captures physical channel mapping or compensates for other differences between domains. For example, when target domain has different class distribution, the class information can change the channel profiles. The optimal transport interpretation does not resolve this ambiguity. Controlled experiments that vary class proportions independently of channel changes might help addressing this concern.
  • I would recommend using EEG datasets (e.g., SEED, DREAMER, BCI Competition, etc.) for experiments beyond HAR and sEMG. EEG would be a valuable additional test, as its channels reflect activity from different brain regions and are very sensitive to electrode placement. Such an evaluation would help clarify whether the method generalizes beyond the rotational or positional changes.
  • The experiments evaluate adaptation within individual datasets, but do not test cross-dataset transfer. Experiments between datasets with comparable sensors, matched channel counts, and shared classes would better support the method's practical value.
Questions for authors

I wrote all the concerns in the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1.

We thank the reviewer for the constructive feedback. We address each point below.

Critical Weakness 1: Importantly, it is unclear whether the learned matrix captures physical channel mapping or compensates for other differences between domains. For example, when target domain has different class distribution, the class information can change the channel profiles. The optimal transport interpretation does not resolve this ambiguity. Controlled experiments that vary class proportions independently of channel changes might help addressing this concern.

In our main evaluation, we include the WISDM dataset, which exhibits natural class imbalance across subjects, as newly discussed in Appendix A.7.1. Despite this inherent imbalance, IDSA demonstrates performance improvements, suggesting that mitigating sensor-wise domain shifts remains beneficial under naturally occurring class-distribution differences. To further investigate how the degree of class-distribution discrepancy affects IDSA, we conduct an additional controlled experiment using the SD-gesture dataset.

Specifically, we construct synthetic binary classification settings by sampling 100 instances from each domain. For the source domain, the class proportions are set to r:(1-r), while the target domain uses the reversed proportions, (1-r):r. By varying r, we examine how IDSA behaves under different degrees of class-distribution discrepancy. The resulting accuracy (%) is reported below.

SD-gesturerOriginal – no IDSAIDSA
ACON0.586.088.2
ACON0.478.479.6
ACON0.360.661.4
ACON0.231.429.6
SASA0.596.698.6
SASA0.493.895.2
SASA0.379.277.4
SASA0.261.257.8

The results indicate that IDSA achieves comparable performance to the baseline under moderate class-distribution differences (r=0.4), whereas its performance becomes slightly lower under more substantial differences (r=0.3 and r=0.2). One possible explanation is that the mini-batch optimization of Eq. 6 relies on averaged representations to characterize sensor-wise dependencies. When the source and target batches have substantially different class compositions, these representations may be affected by the class-distribution discrepancy, potentially influencing the estimation of sensor-channel relationships.

Taken together, the results suggest that IDSA can be sensitive to substantial class-distribution discrepancies, particularly when the source and target domains exhibit strongly reversed class proportions. Nevertheless, we note that the controlled experiment considers an explicitly constructed scenario with reversed class proportions, which may represent a more challenging setting than naturally occurring class imbalance. The improvements observed on WISDM indicate that the benefits of addressing sensor-wise domain shifts can remain substantial under realistic class-imbalance conditions. In other words, we acknowledge that IDSA may be affected by substantial class-distribution discrepancies and leave further investigation into jointly addressing sensor-wise domain shifts and class-distribution discrepancies for future work.

AuthorsOpenReview ↗

Response to Critical Weakness 2

Critical Weakness 2: I would recommend using EEG datasets (e.g., SEED, DREAMER, BCI Competition, etc.) for experiments beyond HAR and sEMG. EEG would be a valuable additional test, as its channels reflect activity from different brain regions and are very sensitive to electrode placement. Such an evaluation would help clarify whether the method generalizes beyond the rotational or positional changes.

We thank the reviewer for suggesting an additional dataset to further evaluate IDSA's applicability. Following this suggestion, we conduct additional experiments on the BCI-IV-2a dataset under a cross-session evaluation setting. The results are reported below.

MethodMean
ACON54.04
IDSA + ACON55.05
SASA63.01
IDSA + SASA63.26
Deep CORAL34.47
IDSA + Deep CORAL35.10
CLUDA35.35
IDSA + CLUDA35.48

IDSA has consistent but marginal effect on BCI-IV-2a. When combined with SASA, IDSA increases the average accuracy from 63.01% to 63.26%. Although the improvements are less pronounced than those observed on the HAR and sEMG datasets, these results highlight the applicability of IDSA to EEG data and its ability to provide additional performance gains in conjunction with existing adaptation methods.

We attribute this difference to the characteristics of cap-based EEG acquisition, where the cap constrains relative electrode positions and may help preserve correspondence between identically indexed source and target channels across sessions. The sensor-wise domain shifts targeted by IDSA may therefore be less pronounced in this setting, and the ideal sensor correspondence matrix Ast\textbf{A}_{st} may be closer to the identity matrix. Under such conditions, explicitly learning sensor-wise alignment provides limited additional benefits and could introduce unnecessary variations into the learned representations, depending on the backbone.

These additional experiments help clarify the scope of IDSA: it is primarily designed to address sensor-wise domain shifts arising from variations in sensor placement and configuration, which are particularly relevant to wearable MTS applications. Its benefits may therefore be less pronounced in controlled acquisition settings where sensor correspondence is largely preserved. We incorporated these findings into the revised manuscript and further clarified the applicability and potential limitations of IDSA across different data acquisition settings.

AuthorsOpenReview ↗

Response to Critical Weakness 3.

Critical Weakness 3: The experiments evaluate adaptation within individual datasets, but do not test cross-dataset transfer. Experiments between datasets with comparable sensors, matched channel counts, and shared classes would better support the method's practical value.

Our formulation in Eq.(2) assumes a shared latent representation U\mathbf{U} across source and target domains, such that the domain discrepancy can be primarily characterized by differences in sensor-wise relationships. This assumption is more reasonable for within-dataset domain adaptation, where source and target data are collected under a common acquisition protocol and share the same underlying task. In contrast, it may not readily hold in cross-dataset transfer settings, where differences in acquisition protocols, sensor characteristics, subject populations, class distributions, and task-specific factors can introduce additional discrepancies in the underlying latent representations. Consequently, cross-dataset transfer involves distribution shifts beyond the sensor-wise variations that IDSA is specifically designed to address.

We clarified this scope in the revised manuscript and acknowledge that IDSA may not consistently improve performance under cross-dataset transfer. Further extending IDSA to account for multiple sources of cross-dataset discrepancies is an interesting direction for future investigation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to all Reviewers

We sincerely thank all reviewers for their careful evaluation. Below, we summarize our responses and revisions, with further details in our individual rebuttals.

1. Contribution of learned inter-domain sensor alignment

  • Identity transport control (idgg, uxrk). We replaced the learned transport matrix with an identity matrix while keeping the remaining components unchanged. On Opportunity HAR, accuracy decreased from 81.96% to 73.99% with ACON and from 82.96% to 71.16% with SASA. Similar degradation was observed on Nina5-18, supporting the contribution of learned cross-domain sensor relations.
  • Alternative alignment strategies (uxrk). IDSA outperformed both alternative approaches, dataset-level Hungarian matching and a simple MLP adapter, across all reported comparisons. It supports the effectiveness of learning inter-domain sensor relationships over predefined 1:1 sensor matching and a simple MLP adapter.

2. Alignment capacity and robustness to different transformations

  • Controlled sensor transformations (uxrk). Under channel permutation, IDSA improved accuracy from 37.5% to 92.9%, matching its performance in the aligned setting. Sign flips remained challenging, consistent with the potential limitations of nonnegative transport. We also extended Fig. 2(a) to evaluate IDSA under the original diagnostic shifts. Deep CORAL + IDSA outperformed Deep CORAL in most permutation settings and several sensor placement conditions. Under orientation shifts, it improved over Deep CORAL in four of six cases and showed smaller accuracy variations across rotation angles than both Deep CORAL and SASA, although it underperformed Deep CORAL at the two smallest angles.
  • Cost functions and measurement variations (uxrk). Alternative cost function experiment showed that alignment quality depends on the chosen cost function. In the controlled permutation experiment, the log-spectral cost achieved 89.6% accuracy compared with 56.0% using the original cost. Additional experiments showed substantial improvements under sensor permutation combined with moderate measurement offsets or temporal phase shifts, while severe offsets remained challenging.

3. Class-distribution differences and generalization

  • Class imbalance (Ti8f, uxrk). IDSA improved performance on WISDM despite natural class imbalance across subjects. A synthetic stress test with deliberately reversed source and target class proportions showed that our method consistently outperformed the baselines under mild distribution discrepancies. Although the gains diminished under more extreme imbalances, we additionally showed that the learned transport plan remained stable.
  • Additional EEG evaluation (Ti8f). On BCI-IV-2a cross-session adaptation, adding IDSA improved mean accuracy for all four UDA backbones by 0.13–1.01 percentage points. A possible explanation for these modest performance gains is less pronounced sensor-wise correspondence mismatch under the cap-based acquisition protocol.
  • Cross-dataset transfer (Ti8f). We clarified that IDSA assumes a shared underlying task and latent representation across domains. Cross-dataset transfer may introduce additional discrepancies beyond sensor-wise shifts, which remain outside the current method's primary scope.

4. Experimental reliability and comparative evaluation

  • Reproducibility and uncertainty (uxrk). We clarified that hyperparameters and checkpoints are selected using source validation data only, without accessing target labels. We expanded Appendix A.7 with dataset preprocessing and split-construction details and conducted five-seed HAR experiments, reporting mean accuracy and standard deviation for the evaluated methods with and without IDSA.
  • Foundation model comparison (uxrk). Our original comparison assessed pretrained FM encoders with source supervision against UDA methods using unlabeled target observations. To address the difference in target data access, we additionally evaluated FM baselines with CORAL loss using unlabeled target data. The additional access to target features yielded inconsistent gains across models and datasets, while the overall performance trend observed in the original comparison remained unchanged.

On manuscript revisions (Ti8f, idgg, uxrk). We expanded the PCA diagnostic description and corrected Appendix A.5.2 to clarify that cross-domain relations guide the weighted aggregation of target signals. We also updated Fig. 2(a) with additional IDSA results, revised the accompanying discussion in Section 4.2 to clarify its applicability and limitations, added the transport control analysis to Section 6.3, and expanded Appendix A.7 with detailed experimental protocols.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.