acceptodds
Under review as a conference paper at ICLR 2027

CONDITION COVERAGE IN SPARSE TARGET CALIBRATION FOR DOMAIN ADAPTATION

Abstract

We study domain adaptation when target labels are available only at a sparse set of calibration levels, while prediction is required across the full target label space. We show that the observed target distributions do not uniquely determine the target distributions at unseen levels, even when the observed distributions are known exactly. This motivates a distributionally robust formulation that accounts for uncertainty induced by incomplete calibration coverage. Based on the resulting robust bound, we propose CaDRA, a Calibration-aware Distributionally Robust Adaptation method that combines labeled target prediction with balanced source– target alignment. Controlled synthetic experiments and three real-world transfer tasks show consistent performance across different calibration patterns and domain shifts.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Domain adaptation, distributionally robust optimization, sparse target calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DiqmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies domain adaptation under sparse target calibration, where labeled source data cover all condition levels, but labeled target data are available only at a subset of calibration levels. The goal is to achieve accurate predictions across the full target condition range, including levels that are not observed during adaptation. The authors formulate the problem from a distributionally robust optimization perspective and derive a theoretical upper bound on the target risk. Based on this bound, they propose CaDRA, which combines coverage-weighted target regression loss with conditional source–target alignment. The proposed method is evaluated on synthetic experiments and three real-world domain adaptation tasks.

Critical strengths

The problem setup is clearly defined and reasonably motivated. In particular, the distinction between observed and unseen target calibration levels is explicit, making the underlying research problem easy to understand.

Critical weaknesses

W1. Limited Methodological Novelty

The proposed training objective appears somewhat conventional, offering mainly incremental methodological contributions. It combines a weighted target calibration loss with source–target alignment, both of which build on established approaches in domain adaptation [1–3]. In particular, Naram et al. [2] derive weighted empirical source losses and a partial Wasserstein alignment term from theoretical bounds. Although the problem settings and weighting schemes differ, the overall bound-driven construction follows a similar approach.

A key methodological distinction of CaDRA lies in its coverage-aware weights αa\alpha_a and βa\beta_a. However, their contribution is not sufficiently validated. A more detailed analysis, including an ablation against uniform weighting while keeping the remaining objective unchanged, is needed to establish their effectiveness.

W2. Insufficient Justification for the Source-Centered DRO Formulation

The paper constructs ambiguity sets centered at the source conditional distributions. However, the proposed DRO formulation does not appear to sufficiently justify the resulting training objective.

Specifically, under the assumed target-side smoothness condition, an alternative formulation would be to center the ambiguity set at the adapted target distribution of the nearest observed calibration level:

Q : W_1 \left(Q,Q^\phi_{a(u)}\right) \leq L_t |u-a(u)| \right\\}$$ Under the target-side smoothness assumption, the true unseen target distribution lies within this ambiguity set. By the triangle inequality, this set is also contained within the proposed source-centered ambiguity set, potentially yielding a tighter robust risk bound. This raises the question of whether the additional source–target alignment term arises primarily from the choice of source-centered ambiguity sets, rather than being inherently necessary under the stated assumptions. Empirically, Table 2 shows only a marginal improvement from Calibration Only 1.670 $\pm$ 0.026 to CaDRA 1.660 $\pm$ 0.023. The authors should provide a clearer theoretical justification for choosing the source-centered formulation over this target-centered alternative, along with stronger empirical evidence demonstrating the additional benefits of the resulting alignment term. **W3. Unclear Experimental Protocol and Baseline Implementation** Several aspects of the experimental protocol remain unclear. Appendix H explicitly mentions the use of target regression loss for CORAL, but does not specify whether the other baselines incorporate the available target labels into their training objectives. This is particularly important because how the labeled data from observed target calibration levels are utilized can substantially affect adaptation performance. The hyperparameter selection procedures also appear asymmetric. While CaDRA's alignment coefficient is selected using validation data, those for CORAL, DANN, and RSD are fixed across experiments, raising concerns about whether comparable tuning was performed. Furthermore, the validation metric used for early stopping and whether the same criteria were consistently applied across baselines remain unclear. Training details are provided primarily for MotorDA, with comparable configurations and early-stopping procedures insufficiently specified for CWRU and Paderborn. The baseline alignment procedures require further clarification. In particular, for DANN and other baselines involving marginal alignment, it is unclear whether source samples from all condition levels were aligned with target samples restricted to the observed calibration levels, or whether the source samples were similarly restricted. Such a configuration could introduce label-distribution mismatch and adversely affect methods such as DANN. Notably, all four baselines underperform Source Only in the Paderborn poor-coverage setting, where only a single target calibration level is available (Figure 3). The authors should provide sufficient implementation details to ensure reproducibility and allow a fair assessment of the baseline comparisons. Without these details, it is difficult to determine whether the reported performance differences reflect genuine methodological advantages or differences in implementation choices. **W4. Insufficient Comparison with Relevant Baselines** The experiments lack several important baselines for the proposed setting. Given the availability of labeled source data and labeled target data at observed calibration levels, joint source–target ERM is a natural supervised baseline. For example, a shared predictor could be trained by minimizing $$\min_{\theta} (1-\eta)\widehat R_s(f_\theta) +\eta\widehat R_{t,\mathcal A}(f_\theta), \qquad \eta\in[0,1],$$ where $\widehat R_{t,\mathcal A}$ is the empirical target regression loss restricted to observed calibration levels. Both the jointly trained predictor and a variant with an additional calibration-only target adapter could be evaluated. These comparisons would help determine whether CaDRA provides benefits beyond straightforward supervised learning with the available labeled data. Supervised and few-shot domain adaptation methods for regression [4, 5], as well as approaches addressing incomplete target-domain coverage [6], are also relevant. Although the latter differ in their target-data assumptions, the authors could discuss whether such methods can be adapted to the proposed setting. The study primarily considers existing alignment-based methods, including CORAL, DANN, RSD, and COD. Broader comparisons would strengthen the empirical evidence supporting CaDRA. Furthermore, the paper reports only overall target MAE. Reporting performance separately for seen and unseen target levels would help distinguish adaptation to observed conditions from generalization to unobserved ones, which is the central goal of this work. **Minor Comments** 1. (Incomplete Related Work) The Related Work section discusses domain adaptation, sparse supervised adaptation, and DRO largely in isolation, without sufficiently addressing research at their intersections. This may give the impression that applying DRO to domain adaptation is itself a novel contribution, although distributionally robust domain adaptation has already been studied [7, 8]. Furthermore, prior work has investigated domain adaptation under incomplete target-domain coverage, such as Li et al. [6]. Although their problem setting and target-data assumptions differ from those of CaDRA, these connections deserve further discussion. I recommend that the authors discuss and cite these relevant studies to better clarify the novelty and positioning of their work. 2. (Ambiguous Axis Labeling in Figure 1) In Figures 1(c) and 1(d), the y-axes are labeled as "Target shift $\delta(c)$." However, $\delta(c)$ is defined as a vector-valued quantity, whereas the plots display its individual scalar components, $\delta_1(c)$ and $\delta_2(c)$. Why is the y-axis labeled with the vector $\delta(c)$ rather than explicitly indicating that scalar components are being plotted? I recommend revising the axis labels to avoid this ambiguity. 3. (Unclear Experimental Settings and Presentation in Figure 3) Figure 3 presents one of the main experimental results, yet its settings are described only as "good coverage" and "poor coverage." These are author-defined rather than standardized evaluation categories, but no precise definition is provided in the main text or figure caption. This is particularly problematic because the calibration settings directly affect the difficulty of the adaptation task. For instance, the Paderborn poor-coverage setting provides labeled target data at only a single level, which is essential for interpreting the substantial performance degradation of the baselines. Should readers have to search through the appendix simply to understand the experimental conditions underlying such a central result? This lack of clarity undermines the interpretability and readability of Figure 3. Moreover, the color scheme appears unnecessarily complex, and the figure omits standard deviations or other measures of variability despite having sufficient space. I recommend explicitly providing the corresponding calibration sets and coverage values in the main text or figure caption, simplifying the visual presentation, and reporting the variability of the results. Overall, given the limited methodological novelty, insufficient justification for the proposed DRO formulation, and concerns about the experimental comparisons, I recommend a clear rejection of this paper in its current form. --- **References** [1] Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gianfranco Doretto. Unified Deep Supervised Domain Adaptation and Generalization. ICCV, 2017. [2] Jayadev Naram, Fredrik Hellström, Ziming Wang, Rebecka Jörnsten, and Giuseppe Durisi. Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport. ICML, 2025. [3] Okan Koç, Alexander Soen, Shanglin Li, and Masashi Sugiyama. Incorporating Importance Weighting in Optimal Transport Based Domain Alignment. ICML, 2026. [4] Takeshi Teshima, Issei Sato, and Masashi Sugiyama. Few-shot Domain Adaptation by Causal Mechanism Transfer. ICML, 2020. [5] Ori Katz, Ronen Talmon, and Uri Shaham. Supervised Domain Adaptation Based on Marginal and Conditional Distributions Alignment. TMLR, 2024. [6] Xiang Li, Wei Zhang, Xu Li, and Hongshen Hao. Partial Domain Adaptation in Remaining Useful Life Prediction With Incomplete Target Data. IEEE/ASME Transactions on Mechatronics, 29(3):1903–1913, 2024. [7] Akram S. Awad and George K. Atia. Distributionally Robust Domain Adaptation. MLSP, 2023. [8] Seonghwi Kim, Sungho Jo, Wooseok Ha, and Minwoo Chae. Distributionally Robust Classification for Multi-source Unsupervised Domain Adaptation. ICLR, 2026.
Questions for authors

Please see the weaknesses.

3 comments

We thank the reviewer for the constructive feedback.

W1. Goal and methodological novelty

We first identified a practical challenge in industrial condition monitoring. A regression model trained on one machine often needs to be adapted to a new machine, while labeled target data can be collected at only a few operating or severity levels. As a result, part of the target range remains unseen during adaptation. We formalize this setting as sparse target-level calibration.

Based on this problem formulation, we propose CaDRA. We show why sparse calibration data alone cannot determine the target distributions at unseen levels without assumptions across levels. We then derive coverage-aware alignment weights from the observed target levels so that the adaptation objective accounts for the parts of the target range that lack calibration data.

Following Reviewer fwWv’s suggestion to use labeled source samples at levels missing from the target calibration data, we revised the objective to Joint-CaDRA, as defined in our response to Reviewer fwWv, Q2. We then evaluated the coverage-aware weighting within this revised objective. The model, data split, target supervision, and selected loss coefficients are fixed. Only the weighting used in LalignL_{\mathrm{align}} is changed.

SettingUniform weightingCoverage weightingBetter
MotorDA-good1.7089 +/- 0.02911.6907 +/- 0.0127Coverage
MotorDA-poor2.3423 +/- 0.03872.3178 +/- 0.0234Coverage
Paderborn-good0.2077 +/- 0.00450.1909 +/- 0.0145Coverage
Paderborn-poor0.2600 +/- 0.01950.2600 +/- 0.0195Same
CWRU-good1.3637 +/- 0.00961.4117 +/- 0.0502Uniform
CWRU-poor5.2290 +/- 0.05145.0887 +/- 0.0179Coverage

Coverage weighting achieves lower overall MAE in four of the six settings and the same MAE in one setting. This ablation directly evaluates the coverage-aware component of the method.

W2. Why the formulation is source-centered

A target-centered ambiguity set is valid when the observed target levels provide sufficient coverage. Under sparse calibration, however, the target domain may contain only one observed anchor. In that case, every unseen target level would use the same target distribution as its reference. The formulation would therefore concentrate all unseen levels around one observed point and provide no level-specific reference for the rest of the target range.

The source domain is fully labeled. Each unseen target level can therefore use the source conditional distribution at the corresponding label as a separate reference. This preserves information across the full label range instead of forcing all unseen levels to depend on a single target anchor. The source-centered formulation is motivated by this practical difference in coverage. We clarified that it is a modeling choice rather than a theoretical necessity, and that a target-centered formulation is a reasonable alternative when sufficient target anchors are available.

W3. Controlled experimental protocol

We reran all methods under the same experimental protocol. Every adaptable method uses the same labeled target calibration samples, target regression loss, frozen source predictor, target-adapter architecture, data split, validation metric, checkpoint rule, and three random seeds. Each adaptation coefficient is selected through a grid-search sweep using validation MAE. No target test label is used for hyperparameter selection or checkpoint selection.

For every alignment-based method, the source samples used in the alignment term are restricted to the observed target labels. CORAL, DANN, RSD, COD, and Joint-CaDRA therefore use the same source samples for alignment.

The table reports the rerun results.

SettingSource OnlyCORALDANNRSDCODJoint-CaDRA
MotorDA-good2.72091.6951 +/- 0.01401.6879 +/- 0.02041.6926 +/- 0.00871.6796 +/- 0.01501.6907 +/- 0.0127
MotorDA-poor2.72092.3700 +/- 0.03332.3949 +/- 0.02812.3741 +/- 0.01692.3851 +/- 0.04922.3178 +/- 0.0234
Paderborn-good0.58280.2038 +/- 0.00560.2358 +/- 0.01970.2213 +/- 0.00970.2067 +/- 0.01760.1909 +/- 0.0145
Paderborn-poor0.58280.5390 +/- 0.06250.8282 +/- 0.01730.4974 +/- 0.04140.5813 +/- 0.05720.2600 +/- 0.0195
CWRU-good6.50421.8488 +/- 0.07202.2705 +/- 0.05621.9558 +/- 0.07751.7610 +/- 0.08281.4117 +/- 0.0502
CWRU-poor6.50425.7431 +/- 0.03465.6048 +/- 0.02845.5756 +/- 0.02555.6934 +/- 0.04905.0887 +/- 0.0179

Joint-CaDRA achieves the lowest overall MAE in five of the six settings and remains close to the best result on MotorDA-good.

W4. Additional baselines and observed-unseen results

We added both requested supervised baselines under the W3 protocol:

  • Calibration Only minimizes LCal=LTL_{\text{Cal}} = L_T (trains the target adapter using only observed target labels).
  • Joint ERM minimizes LJointERM=LT+λsLSL_{\text{JointERM}} = L_T + \lambda_s L_S (uses observed target labels and labeled source samples, without alignment loss).

Following Reviewer fwWv's suggestion to use source labels at levels missing from target calibration data, we revised CaDRA to Joint-CaDRA: LJointCaDRA=LT+λsLS+λaLalignL_{\text{JointCaDRA}} = L_T + \lambda_s L_S + \lambda_a L_{\text{align}} which adds the original CaDRA alignment loss to Joint ERM. We also evaluated DARE-GRAM, designed for continuous-output domain adaptation.

SettingSource OnlyCalibration OnlyJoint ERMDARE-GRAMJoint-CaDRA
MotorDA-good2.72091.6837 ± 0.01411.9421 ± 0.03051.6660 ± 0.02601.6907 ± 0.0127
MotorDA-poor2.72092.3772 ± 0.06452.2901 ± 0.00462.3790 ± 0.03802.3178 ± 0.0234
Paderborn-good0.58280.2237 ± 0.01280.2739 ± 0.00100.2330 ± 0.00200.1909 ± 0.0145
Paderborn-poor0.58280.5393 ± 0.08990.3051 ± 0.00570.4090 ± 0.05400.2600 ± 0.0195
CWRU-good6.50422.0568 ± 0.06121.4519 ± 0.00701.4710 ± 0.06301.4117 ± 0.0502
CWRU-poor6.50425.5821 ± 0.01285.2057 ± 0.00915.3150 ± 0.12605.0887 ± 0.0179

Joint-CaDRA outperforms Calibration Only, Joint ERM, and DARE-GRAM in 5/6 comparisons each, achieving the lowest overall MAE in 4 settings and second-lowest in the other 2.

Below we separate performance on observed vs. unseen target levels:

SettingMethodObserved MAEUnseen MAEOverall MAE
MotorDA-goodSource Only3.4571.6162.721
MotorDA-goodJoint-CaDRA1.3932.1381.691
MotorDA-poorSource Only2.7912.6152.721
MotorDA-poorJoint-CaDRA1.3083.8322.318
Paderborn-goodSource Only0.6410.4670.583
Paderborn-goodJoint-CaDRA0.0700.4340.191
Paderborn-poorSource Only1.0870.3310.583
Paderborn-poorJoint-CaDRA0.0240.3780.260
CWRU-goodSource Only7.5053.5036.504
CWRU-goodJoint-CaDRA0.3494.6001.412
CWRU-poorSource Only2.36210.6466.504
CWRU-poorJoint-CaDRA0.3429.8365.089

Joint-CaDRA reduces unseen MAE on Paderborn-good and CWRU-poor, while negative transfer remains in the other four settings.

We added Causal Mechanism Transfer [1] and partial domain adaptation with incomplete target data [2] to Related Work. [1] assumes multiple labeled source domains to learn a shared generative mechanism, whereas our setting starts from a single fixed pretrained source predictor without retraining it. [2] studies unlabeled partial degradation trajectories for remaining useful life prediction, whereas we use a few labeled target levels and evaluate on unseen levels. Because [1, 2] require different data and training access, we directly evaluated Calibration Only, Joint ERM, and DARE-GRAM under the same frozen-source, target-calibration protocol.

[1] T. Teshima et al., "Few-shot Domain Adaptation by Causal Mechanism Transfer," ICML, 2020. [2] X. Li et al., "Partial Domain Adaptation in Remaining Useful Life Prediction With Incomplete Target Data," IEEE/ASME Trans. Mechatronics, 2024.

Minor comments and presentation corrections

We addressed all minor comments directly:

  • Defined the W1W_1 ground metric as the unsquared Euclidean norm, making W1≤W2W_1 \le W_2 in Eq. (10) explicit.
  • Added the missing space before =1/K= 1/K.
  • Changed Fig. 3's y-axis label from "Improvement (%)" to "Reduction in Target MAE (%)".
  • Replaced the capitalized coupling symbol in Appendix E with a lowercase symbol.
  • Added a discussion of negative transfer under poor target coverage.
  • Relabeled the scalar shift components in Fig. 1(c,d).
  • Added observed anchor sets, unseen sets, and numerical coverage values to the main text.
  • Replaced percentage-only results with absolute observed, unseen, and overall MAE values and standard deviations.
  • Simplified figure colors and standardized method names.
  • Added the Paderborn bearing-to-level mapping and complete CWRU/Paderborn architectures.
  • Specified the optimizer, batch construction, search grids, validation metric, and checkpoint rule for all methods.
  • Replaced "paired windows" with "independently extracted windows from each domain" (no sample-level correspondence).

I thank the authors for their efforts in revising the manuscript and providing additional experiments. Nevertheless, I believe that several substantial concerns remain unresolved. More generally, I find that the response does not directly address several of the specific points raised in my original review.

First of all, the authors introduce a revised training objective, Joint-CaDRA, without providing a sufficiently clear definition, motivation, or justification in their response. Instead, they refer me to their response to another reviewer. I do not think it is reasonable to expect a reviewer to locate and interpret another reviewer's response simply to understand a major methodological change that is presented as part of the rebuttal to their own concerns.

Methodological Novelty. The authors reiterate the motivation for their problem setting and provide additional experiments, but do not directly explain the methodological differences from the prior works identified in my review. Moreover, the ablation only partially evaluates the proposed coverage-aware weights. My concern regarding methodological novelty therefore remains unresolved.

Source-Centered DRO Formulation. The response does not address my mathematical argument that the target-centered ambiguity set remains valid under sparse calibration and is contained within the proposed source-centered set. The availability of level-specific source distributions does not establish the theoretical justification of the source-centered formulation or its alignment term.

Minor Comments. Although some references have been added, the authors do not provide substantive comparisons with relevant prior work, particularly existing DRO-based domain adaptation studies, as requested. Such comparisons are important for assessing methodological novelty. Other minor comments also remain insufficiently addressed.

I therefore maintain my original assessment.

Reviewer_7BH5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on a specific setup where we want to perform domain adaptation but where we lack data for some levels in our new adapted domain. It is quite a specific problem compared to having little data on every label, but it is clearly relevant and useful, especially for some industrial applications. The proposed method (CaDRA) trains a small adapter on top of a frozen source model, using the available target data while giving more focus to the levels close to the missing ones, and aligning target and source features matched by operating conditions. The experiments use synthetic data and three real-world datasets, compared against CORAL, DANN, RSD and COD.

Critical strengths

The paper is well motivated and the tackled problem is clearly introduced. The proposed method obtains better performance than the the SOA methods it is compared to.

Critical weaknesses

Theory:

The formulation of the problem seems overly complex and Theorem 1, described as the first contribution, seems too straightforward, as it follows directly from the Lipschitz assumption, to be a real contribution.

Experiments:

The results shown in the experiments only represent MAE averaged across all levels. We thus don’t have information about the performance on the levels where we have data or not. The paper's motivation is to have better performance especially on those levels so it should be shown in the experiments.

For the real-world data we only have percent improvement where the proposed method only shows a low gain over the baselines for half of the setups.

No real ablation has been done on the weights ɑ and β (e.g. compared to uniform weights), although they are the main element coming from the theory that is used in the final objective. Moreover, the only ablation shows that calibration only obtains almost the same results as CaDRA, which questions the benefit of the alignment term and method.

Most baselines (DANN, RSD, COD) are unsupervised DA methods not designed to use target labels, part of CaDRA's gain may come from the label information rather than from the proposed alignment.

Questions for authors

What are the consequences of the assumptions made in 4.1 for real-world use cases? Why are you using only the nearest labeled data and not the available data around it turning the extrapolation issue into an interpolation one? Any intuition about the huge performance gain on the Paderborn dataset (and CWRU poor coveraged) compared to the other small gain?

Can you report MAE separately on observed and unseen levels?

1 comment

We thank the reviewer for the thoughtful questions, which helped us clarify the role of the theory, the use of target-level coverage, and the interpretation of the results.

1. Role of Theorem 1

Theorem 1 formalizes a key observation rather than emphasizing technical complexity. It shows that sparse target calibration cannot determine the target distributions at unseen levels without additional cross-level assumptions. This observation motivates the subsequent coverage-aware formulation and training objective. We revised the contribution statement to make this role clear.

2. Performance on observed and unseen target levels

We reported MAE separately at the observed calibration levels and unseen target levels. Observed MAE measures performance at levels with target labels, while unseen MAE measures performance at levels without target labels. As suggested by Reviewer fwWv, we revised our objective to Joint-CaDRA to reduce negative transfer at unseen levels. Its objective and overall results are provided in our response to Reviewer fwWv, Q2.

SettingMethodObservedUnseenOverall
MotorDA-goodSource Only3.4571.6162.721
Joint-CaDRA1.3932.1381.691
MotorDA-poorSource Only2.7912.6152.721
Joint-CaDRA1.3083.8322.318
Paderborn-goodSource Only0.6410.4670.583
Joint-CaDRA0.0700.4340.191
Paderborn-poorSource Only1.0870.3310.583
Joint-CaDRA0.0240.3780.260
CWRU-goodSource Only7.5053.5036.504
Joint-CaDRA0.3494.6001.412
CWRU-poorSource Only2.36210.6466.504
Joint-CaDRA0.3429.8365.089

Joint-CaDRA achieves lower overall MAE in all six settings, mainly due to the large reduction at observed levels. It also reduces unseen MAE in two settings, although negative transfer remains in the other four.

3. Why the nearest observed level is used

The nearest observed level gives the smallest distance term in the bound. Using observations from both sides may be useful when the unseen level lies between two labeled levels. However, this requires labeled data on both sides and an additional assumption about how the target distribution changes between them. Neither condition is guaranteed under sparse target coverage. In particular, when the available labels lie only on one side of the unseen level, interpolation is not possible. We therefore use the nearest observed level as a consistent and conservative choice under the current assumptions.

4. Contribution of coverage-aware alignment

All methods use the same labeled target calibration samples and the same target regression loss. They also use the same data split, validation procedure, and model-selection criterion. The comparison is therefore made under the same target supervision.

To isolate the contribution of coverage-aware alignment, we keep the model, data split, and selected loss coefficients fixed and change only the weighting used in LalignL_{\mathrm{align}}. The table reports overall MAE over three random seeds.

SettingUniform weightingCoverage weightingBetter
MotorDA-good1.7089 +/- 0.02911.6907 +/- 0.0127Coverage
MotorDA-poor2.3423 +/- 0.03872.3178 +/- 0.0234Coverage
Paderborn-good0.2077 +/- 0.00450.1909 +/- 0.0145Coverage
Paderborn-poor0.2600 +/- 0.01950.2600 +/- 0.0195Same
CWRU-good1.3637 +/- 0.00961.4117 +/- 0.0502Uniform
CWRU-poor5.2290 +/- 0.05145.0887 +/- 0.0179Coverage

Coverage weighting achieves lower overall MAE in four of the six settings and the same MAE in one setting. This ablation shows that coverage-aware weighting provides an additional improvement in most evaluated settings.

5. Why the gains vary across datasets and coverage settings

The gains vary because the datasets and coverage settings are different. In some settings, the unseen levels are close to the observed target levels and are better supported by the available calibration data. In others, more of the target range is unseen or farther from the observed levels, making prediction more difficult. The resulting performance therefore differs across datasets and coverage settings.

Reviewer_fwWvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies domain adaptation for regression when labeled target data are available only at a subset of label levels. The authors establish a non-identifiability result for unseen target distributions and derive a distributionally robust upper bound on the full-target prediction risk. Based on this bound, they propose CaDRA, which combines supervised target calibration with source–target alignment. Experiments are conducted on synthetic and three real-world datasets.

Critical strengths
  1. The paper considers a practically relevant setting where entire target label levels are unavailable during adaptation, rather than simply having a limited number of labeled target samples.

  2. The paper provides a theoretical motivation for the proposed objective and evaluates the method under different domain shifts and calibration patterns.

Critical weaknesses

1. Strong assumptions and concerns about the problem formulation.

The theoretical results rely heavily on the cross-level Wasserstein smoothness assumptions in Eq. (3). These assumptions are imposed on the latent distributions, which depend on the feature extractor and, for the target domain, the adapter. In particular, the target-side condition is required to hold for all considered adapters ϕ\phi. I find this assumption rather strong, especially since the proposed objective does not explicitly enforce such smoothness. It is also unclear how these assumptions can be justified or verified in the real-world experiments.

2. Concerns about the upper bound and the resulting training objective.

The proposed training objective is derived from Theorem 2, but I am concerned that the upper bound may be too loose to provide useful guidance for adaptation, especially when calibration coverage is highly asymmetric.

For example, consider C={1,…,10}\mathcal C=\{1,\ldots,10\} and A={1,2,3}\mathcal A=\{1,2,3\}, with uniform deployment weights. All unseen levels are assigned to calibration level 3, giving α3=0.8\alpha_3=0.8 and β3=0.7\beta_3=0.7, while α1=α2=0.1\alpha_1=\alpha_2=0.1 and β1=β2=0\beta_1=\beta_2=0. As a result, the supervised loss is dominated by level 3, and the alignment loss completely ignores levels 1 and 2. I am concerned that such imbalanced weights may cause the model to overfit a particular calibration level and hurt generalization to unseen levels.

This issue seems closely related to how the upper bound is constructed. The proof controls the target risk at an unseen level u through its nearest calibration level a(u), using the source distribution as an intermediate step. However, the bound does not directly use the source predictor's performance at level u, even though the source model is trained on the full label space. In the example above, levels 4–10 are all associated with level 3, regardless of how far away they are. Their distances only contribute to the coverage term, which is independent of ϕ\phi and therefore does not affect optimization.

This may be particularly problematic when the source–target discrepancy is small. In that case, the pretrained source model may already perform well at unseen target levels, but the proposed objective does not explicitly preserve its predictive behavior on those levels. The adaptation process could therefore introduce negative transfer.

Overall, I am not convinced that minimizing this upper bound necessarily leads to a good adaptation objective. The authors should better justify the coverage-derived weights, particularly under highly asymmetric calibration designs, and investigate whether the proposed adaptation can degrade the source model's performance at unseen levels.

Questions for authors
  1. Operating variable M. Why is M used by the target adapter but not by the pretrained source model, given that it is observed in both domains? In particular, Line 121 defines Zs=Es(X)Z_s=E_s(X) for (X,M,C)∼Ps(X,M,C)\sim P_s, but M does not enter the source representation.

  2. Baseline comparisons. Most baselines were originally proposed for unsupervised domain adaptation, while the current setting provides labeled target calibration data. Appendix H describes the adaptations, but does not clearly specify whether all baselines use the same supervised target loss. Could the authors clarify their training objectives and provide a controlled comparison using the same supervised loss?

  3. Smoothness assumptions. Under what conditions would Eq. (3) hold for the learned representations, particularly for all considered target adapters? Is there empirical evidence supporting these assumptions on the real-world datasets?

1 comment

We thank the reviewer for the constructive feedback.

1. Cross-level smoothness assumption

We agree that Eq. (3) is an idealized modeling assumption rather than a property guaranteed by the model. Our theoretical analysis is derived under this assumption, and we do not claim that it holds for every dataset or adapter in practice. Nevertheless, the method derived from this theoretical setting achieves competitive empirical performance across multiple real-world datasets and coverage settings. We revised the paper to distinguish the conditional theoretical analysis from the empirical effectiveness of the method.

2. Upper bound and the training objective

We agree that the original upper bound can be loose under asymmetric calibration coverage. Our observed and unseen evaluation confirms that negative transfer can occur. On CWRU-good, the original objective reduces overall MAE from 6.504 to 1.346, while increasing unseen-level MAE from 3.503 to 3.938.

Motivated by the reviewer’s suggestion to use labeled source samples at levels missing from the target calibration data, we revised the objective to

$ L_{\mathrm{JointCaDRA}} = L_T + \lambda_s L_S + \lambda_a L_{\mathrm{align}}. $

Here, LTL_T uses the observed target labels, LSL_S uses labeled source samples across the complete label range, and LalignL_{\mathrm{align}} is the original alignment loss. The positive coefficients λs\lambda_s and λa\lambda_a are selected using validation data without target test labels.

We evaluated the revised objective on all six real-world settings.

SettingSource OnlyCORALDANNRSDCODJoint-CaDRA
MotorDA-good2.72091.6951 +/- 0.01401.6879 +/- 0.02041.6926 +/- 0.00871.6796 +/- 0.01501.6907 +/- 0.0127
MotorDA-poor2.72092.3700 +/- 0.03332.3949 +/- 0.02812.3741 +/- 0.01692.3851 +/- 0.04922.3178 +/- 0.0234
Paderborn-good0.58280.2038 +/- 0.00560.2358 +/- 0.01970.2213 +/- 0.00970.2067 +/- 0.01760.1909 +/- 0.0145
Paderborn-poor0.58280.5390 +/- 0.06250.8282 +/- 0.01730.4974 +/- 0.04140.5813 +/- 0.05720.2600 +/- 0.0195
CWRU-good6.50421.8488 +/- 0.07202.2705 +/- 0.05621.9558 +/- 0.07751.7610 +/- 0.08281.4117 +/- 0.0502
CWRU-poor6.50425.7431 +/- 0.03465.6048 +/- 0.02845.5756 +/- 0.02555.6934 +/- 0.04905.0887 +/- 0.0179

Joint-CaDRA achieves the lowest overall MAE in five of the six settings and remains close to the best result on MotorDA-good.

We also tested whether the proposed coverage weights improve the alignment term in Joint-CaDRA. The model, data split, and selected values of λs\lambda_s and λa\lambda_a are fixed. The only change is the weighting used in LalignL_{\mathrm{align}}. Uniform weighting assigns the same weight to every observed target level. Coverage weighting assigns each observed target level a weight based on the unseen levels for which it is the nearest available anchor.

SettingUniform weightingCoverage weightingBetter
MotorDA-good1.7089 +/- 0.02911.6907 +/- 0.0127Coverage
MotorDA-poor2.3423 +/- 0.03872.3178 +/- 0.0234Coverage
Paderborn-good0.2077 +/- 0.00450.1909 +/- 0.0145Coverage
Paderborn-poor0.2600 +/- 0.01950.2600 +/- 0.0195Same
CWRU-good1.3637 +/- 0.00961.4117 +/- 0.0502Uniform
CWRU-poor5.2290 +/- 0.05145.0887 +/- 0.0179Coverage

Coverage weighting achieves lower overall MAE in four of the six settings and the same MAE in one setting.

3. Use of the operating variable MM

CaDRA assumes access to a source encoder gsg_s that was trained and provided by the source-domain model owner before target adaptation. The encoder maps the feature vector XX to z=gs(X)\mathbf z=g_s(X). Its parameters remain fixed, and CaDRA begins from the resulting representation z\mathbf z. The target adapter takes both z\mathbf z and the available operating metadata MM, allowing it to transform the representation differently at different speeds, loads, or torques. The coupling also uses MM to match target samples with source samples collected under similar operating conditions.

4. Baseline objectives and fair tuning

All adaptable baselines use the same labeled target calibration samples and the same target regression loss LTL_T. They also use the same frozen source predictor, target adapter, data split, and validation procedure. The methods differ only in the adaptation term, with Lmethod=LT+λLadaptL_{\mathrm{method}}=L_T+\lambda L_{\mathrm{adapt}}.

For CORAL, DANN, RSD, COD, and CaDRA, LadaptL_{\mathrm{adapt}} is replaced by the corresponding method-specific objective. Every method therefore receives the same target supervision. We restated these objectives and the common training protocol explicitly in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.