acceptodds
Withdrawn from ICLR 2027

Uncertainty-Calibrated Label Density Regularization for Few-Labeled Source-Free Regression Domain Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-labeled source-free regression adapts a pretrained regressor using scarce labeled target data and abundant unlabeled target inputs, without accessing source samples. Observed responses provide verified but sparse supervision, whereas predictions offer broader coverage but may be biased under domain shift. We propose Uncertainty-Calibrated Label Density Regularization (UCLDR), which uses predictions as uncertainty-regulated response evidence rather than directly fitting them as hard point-wise pseudo-labels. Uncertainty controls the admission, probability mass, and spread of predicted responses. The mixture supplies global mean–variance constraints, while its constituent assignments connect unlabeled features to labeled prototypes in overlapping response regions. This shared representation supports both forms of supervision without a separate trainable density model. At ten target labels, UCLDR achieves the lowest mean MSE on three of four benchmarks and the best average rank of 1.25 among the evaluated adaptation methods. On Cycle, it reduces mean MSE relative to CRAFT by 3.45–10.36% across five label budgets, with lower error on 7 to 11 of 12 transfers at each budget.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Source-free domain adaptation, Domain adaptation regression, Few-labeled target adaptation, Uncertainty estimation, Label density, Evidential regression regularization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XhYUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies regression adaptation with a pretrained source model, a small labeled target set, and unlabeled target inputs. Its proposed UCLDR method constructs a discrete response mixture from observed labels and model predictions. Evidential predictive mean uncertainty controls which predictions enter this mixture, their relative masses, and their kernel widths. The mixture supplies mean and variance targets for output regularization, while its individual response assignments support alignment between unlabeled feature prototypes and labeled feature prototypes. Additional evidence and response penalties regulate optimization, and the representation is periodically refreshed with stopped gradients. Experiments cover four tabular regression benchmarks with twelve directed transfers each. At ten target labels, UCLDR has the best average adaptation rank and leads on three benchmarks. Additional Cycle experiments examine five label budgets. Appendices provide component comparisons, uncertainty replacements, density fidelity measurements, optimization diagnostics, and implementation settings. The results suggest useful local supervision, but do not establish the necessity or reliability of the complete density construction.

Critical strengths
  1. The method explicitly separates the mass of prediction evidence from its kernel spread through per kernel normalization in Equations 4 to 6. The shared assignments provide a coherent way to connect marginal output statistics with labeled feature prototypes without introducing another trainable density network.

  2. The evaluation includes all twelve directed transfers on each benchmark and multiple label budgets on Cycle. The appendices also disclose adverse uncertainty results, labeled only fallbacks, and large failures of partial objectives, which make the method's limitations easier to assess.

Critical weaknesses
  1. The comparative protocol is insufficiently specified. Section 4.1 and Appendix C describe UCLDR preprocessing and selection, but not matching baseline procedures for source training, target splits, label access, or response coordinates. Table 2 reports Housing MSE of 6.997 for fully supervised Target Only versus 1.449 for UCLDR, alongside severe Cycle baseline failures. These outcomes are possible, but require explanation and matched protocols before attributing the advantage to UCLDR. Excluding test labels from its selection alone does not establish comparability.

  2. Density regularization's contribution is not isolated. Tables 4 and 5 retain density regularization in every ablation and omit matched supervised adaptation or supervised adaptation with only local alignment. On Cycle at twenty labels, density and safeguarded density yield MSEs of 474.140 and 331.086, versus 1.049 for the complete method. These large failures indicate instability in the global pathway. Without removing it from the successful configuration, the experiments cannot distinguish useful density supervision from local supervision rescuing an unnecessary or harmful objective.

  3. Uncertainty regulation can fail or become ineffective. Tables 6 and 7 show that random admission and fixed bandwidth improve Kin8xy, where uncertainty correlates negatively with error and selected predictions are worse. Appendix E reports all 144 canonical Housing checkpoints retaining the full pool with uncertainty at its ceiling and essentially uniform weights. Although mass allocation helps two datasets, Appendix C specifies different learning rates and normalization for these controls. They therefore do not establish that the coordinated uncertainty mechanism explains the main benchmark gains.

  4. The mixture need not estimate the target population moments. Equations 3 to 6 select and reweight predictions by uncertainty, assigning them half the mixture mass regardless of coverage. Even accurate predictions then represent a reliability weighted marginal, not the target marginal. Kernel broadening also adds dispersion that Equation (8) asks predictive means to reproduce. Appendix B.1 does not control these biases, and Table 8 lacks label only or unweighted comparators. Mass conservation therefore does not justify the enforced mean and variance.

  5. Prototype alignment does not establish the required response association. Equations 9 to 11 align feature means after grouping inputs by cached predictions. Incorrect assignments can reinforce incorrect responses, while individual errors can cancel within prototype averages. Scarce labels also leave response regions unsupported. Appendix B.2 assumes the positive prediction and response association required by the risk bound rather than deriving it from alignment. The local branch helps empirically on Cycle, but no association diagnostic establishes when it corrects rather than reinforces errors.

Questions for authors

See Weaknesses.

Reviewer_DJVmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Uncertainty-Calibrated Label Density Regularization (UCLDR) for few-labeled source-free regression domain adaptation. The method uses predictive uncertainty to control the selection, weighting, and distribution of unlabeled predictions, constructing a shared response representation. It combines global mean-variance regularization with local prototype alignment to improve adaptation without directly relying on hard pseudo-labels. Experiments on four regression benchmarks (Bike, Cycle, Housing, and Kin8xy) demonstrate competitive performance, achieving the best average rank among the evaluated adaptation methods. Additional ablation studies and uncertainty analyses support the effectiveness of the proposed framework.

Critical strengths
  1. Meaningful approach to uncertainty-aware regression adaptation. The paper addresses the important challenge of learning from scarce labeled target data while avoiding unreliable pseudo-label supervision. Combining uncertainty-regulated response distributions with global and local supervision provides a reasonable approach to improving source-free regression adaptation.

  2. Promising experimental results. UCLDR achieves the lowest mean MSE on three of four benchmarks with ten target labels. The improvements across multiple label budgets on Cycle, together with component analyses, provide empirical evidence supporting the proposed method.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework mainly combines established techniques, including uncertainty-based sample selection, kernel-based density estimation, moment matching, and prototype alignment. Although their integration is reasonable, the paper does not sufficiently demonstrate a fundamental methodological advance over existing regression adaptation methods.

  2. Inconsistent improvements and limited experimental generalization. UCLDR does not consistently outperform strong baselines across all benchmarks. For example, on Kin8xy, its MSE (1.419) is substantially higher than CRAFT (0.988). Moreover, the main experiments focus on tabular regression benchmarks, and the multi-budget evaluation is primarily conducted on Cycle. This limits the evidence for the method's robustness and generalizability across different regression tasks and domain shifts.

Questions for authors

Q1. How does UCLDR fundamentally differ from existing uncertainty-weighted pseudo-labeling and distribution-based regression adaptation methods? Could the authors clarify the key methodological novelty beyond combining established techniques?

Q2. Why does UCLDR underperform CRAFT on Kin8xy (1.419 vs. 0.988 MSE)? Could the authors explain the conditions under which uncertainty-regulated response representations become less effective?

Q3. Could the authors provide multi-budget evaluations on Bike, Housing, and Kin8xy, rather than primarily on Cycle, to further demonstrate the method's robustness across different target-label budgets?

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

This paper proposes UCLDR for few-labeled source-free regression domain adaptation. The method constructs an uncertainty-regulated shared response representation that combines global moment constraints with local feature supervision anchored by observed labels. The method is evaluated on extensive benchmarks, with additional analyses across label budgets and component ablations.

Critical strengths

The paper addresses a practically relevant problem: source-free regression adaptation with limited target labels. The framework combines global moment regularization and local feature supervision through a shared response representation, without introducing a separate trainable density model. Experiments cover four benchmarks and multiple target-label budgets.

Critical weaknesses
  1. The argument between research motivation and methodological choices remains insufficient. Introduction identifies the limitations of point-wise pseudo-labels and marginal statistical supervision. However, it does not fully elucidate the specific challenges that existing methods face in exploiting two forms of supervision and explain how UCLDR address those challenges.
  2. Existing ablation experiments have not verified the additional role of global statistical constraints in the complete model. The component comparisons in Appendix D demonstrate the benefit of adding local supervision while keeping density-moment regularization and both safeguards active. However, they do not examine the effect of removing the global moment constraint from the complete model. Consequently, the additional contribution of the global constraint beyond local supervision remains unclear. Given its relevance to the central contribution, the key ablation results should be presented in the main text.
  3. The high prediction errors in some ablation configurations and the role of the safeguards require further explanation. In Table 5, for 20 and 30 labels, the average MSE for the density configuration with the protection mechanism was 331.086 and 303.903, respectively—significantly higher than the 1.049 and 0.931 values for the full UCLDR. Notably, for 30 labels, adding the protection mechanism actually caused the MSE to rise from 46.897 to 303.903. The authors should clarify the causes of these high errors and why the safeguards fail to improve performance in this configuration.
  4. The clarity and consistency of the presentation could be improved. In particular, the correspondence between the modules depicted in Figure 1 and their methodological descriptions in Section 3 should be clarified, with consistent terminology maintained throughout the manuscript.
Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.