acceptodds
Withdrawn from ICLR 2027

PIVOT: Partially-observed Indexing for Variational Online adapTation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Domain adaptation (DA) in real-world applications often unfolds in an online fashion, where data arrives sequentially with limited domain access and imbalanced sampling across domains. For example, in personalized ads prediction, users from different demographic groups (e.g., countries or age cohorts) correspond to distinct domains with highly skewed data availability, while user interests evolve over time. However, existing domain adaptation methods typically assume full domain observability and balanced data access, limiting their applicability to real-world scenarios with online domain shift and data imbalance. To address these challenges, we propose Partially-observed Indexing for Variational Online adapTation (PIVOT), a continual domain adaptation framework designed for partial domain access and inter-domain sample imbalance. Starting from a base model pretrained on historical source and target domains, PIVOT incrementally updates latent domain indices over time using a smoothed reweighting kernel and a replay buffer to ensure stable adaptation. Experiments on both synthetic and real-world datasets demonstrate that PIVOT consistently outperforms state-of-the-art baselines in long-term accuracy under dynamic and resource-constrained conditions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Imbalanced Learning, Domain Adaptation, Probabilistic methods

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_XbCaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online imbalanced domain adaptation with a fixed collection of labeled source domains and unlabeled target domains. All domains evolve over time, but each adaptation round provides fresh data from only a subset, with unequal sample counts. PIVOT starts from a variational domain indexing model trained on balanced historical data covering every domain. It adds a temporal prior that regularizes each current domain index toward its previous distribution, a Gaussian kernel that smooths domain sample counts before inverse weighting, and a replay buffer that supplies historical examples for inaccessible domains. Experiments on Growing Circle, CompCars, and temperature forecasting compare adaptation within the current round and generalization to the next round. Reported aggregate results improve over online VDI, and CompCars ablations examine the three additions. The appendix also includes an adapted continuous domain adaptation baseline. The central claim is that these components enable stable adaptation under simultaneous drift, incomplete domain access, and sampling imbalance.

Critical strengths
  1. The problem formulation captures a useful combination of constraints. Section 3 distinguishes persistent domain identities, temporal changes within each domain, and incomplete access to fresh observations. This makes the setting more informative than a single static transfer problem.

  2. The experimental evidence includes classification and regression, current and subsequent round evaluation, and component removal studies. Tables 1 through 8 show aggregate improvements over online VDI, while Appendix F adds a continuous domain adaptation comparison. These results support the practical promise of the combined method within the tested setting.

Critical weaknesses
  1. The temporal objective is not the ELBO of the claimed product prior. Section 3.2 and Equation (7) replace the KL to a normalized product of Gaussian priors with two weighted KL terms. These differ by a posterior entropy term and a normalization term, so they are not equivalent when posterior variance is learned. Appendix A supplies no correction. The objective may be useful regularization, but its central probabilistic justification is unsupported.

  2. Reweighting depends on arbitrary domain numbering. Equation (4) restricts candidate neighbors to an integer ID window before weighting them by learned index distance. Relabeling domains can therefore change the objective while preserving their geometry. Appendices C and E establish neither a compatible ordering for viewpoints and states nor a procedure for updating neighborhoods. Without geometry based selection or an explicit ordering assumption, the claimed continuous reweighting mechanism lacks a justified definition.

  3. Stale replay does not establish adaptation to missing domains' current distributions. Section 3.2 replaces inaccessible domains with historical samples, while the temporal prior favors their previous indices. There is no drift prediction, age dependent weighting, or uncertainty adjustment during absence. Appendix G acknowledges abrupt shift failures, but the experiments do not separate inaccessible domain performance or test absence duration. Long gaps can accumulate substantial smooth drift, leaving the central partial observation claim unverified.

  4. Baseline controls do not isolate the indexing innovations. Tables 7 and 8 remove components but omit matched comparisons with replay alone, ordinary frequency balancing, and simpler temporal regularization. These ablations show that components help the combination, not that learned index smoothing is necessary. The variant without temporal priors already outperforms online VDI substantially. Appendix F adds CDA, but does not establish matched memory and balancing conditions, leaving the distinctive contribution unresolved.

  5. Evaluation information boundaries are unspecified. Section 4.1 and Appendices C and D.1 do not clearly separate adaptation examples from current round evaluation or identify labeled versus unlabeled domains. Temperature forecasting uses six future months as outputs without explaining how training windows stay within available time blocks. This does not prove leakage, but prevents verifying whether next round scores reflect prediction on unseen data using only information available at the stated time.

Questions for authors

See Weaknesses.

Reviewer_KxbDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper adapts online domain adaptation with timely evolving source and target distributions, partial domain availability, and imbalanced sample counts across domains. To tackle these issues, the authors introduce an extension of Variational Domain Indexing, with incorprating teporal priors, reweighting strategies and a replay buffer to refine latent domain representations. Experimental results on synthetic and real-world datasets are shown in the paper to support the claim.

Critical strengths

The paper is easy to follow.

Critical weaknesses
  • The significance of the proposed problem formulation remains insufficiently established. The author highlight three aspects of the problem setting, 1) Multi-domain Online Evolution, 2) Partial Domains available. 3) Data imbalance. However, closely related challenges have been extensively studied in different terms, adaptive learning under concept drift[1], continual test-time adaptation[3], including settings that do not require domain identities[2] or access to source data during adaptation. The paper does not sufficiently explain what additional technical challenges arise from this combination.

  • Unclear benefit of explicit domain indexing.The paper builds on the premise that explicitly modeling domain identities is essential for effective adaptation. However, few line of work does not explicitly model the domain identities. [7,8].

  • Limited demonstrated impact. It remains unclear whether their integration of the proposed temporal prior[4], reweighting[5], replay buffer[6] is a meaningful practical advantage established adaptation strategies.

  • Lack of Validation. The baselines are compared against offline DA methods, rather than methods developed for related streaming or continual adaptation settings.

  • Lack of Experimental details about models, training method, and procedures.

[1] Online Boosting Adaptive Learning under Concept Drift for Multistream Classification, AAAI 2024 [2] Open Compound Domain Adaptation, CVPR 2020 [3] Continual Test-Time Domain Adaptation CVPR 2022

[4] Generalizing to Evolving Domains with Latent Structure-Aware Sequential Autoencoder,ICML 2022 [5] Dynamic Weighted Learning for Unsupervised Domain Adaptation, CVPR 2021 [6] ConDA: Continual Unsupervised Domain Adaptation, 2021

[7] Efficient Test-Time Adaptation of Vision-Language Models, CVPR 2024, [8] BATCLIP: Bimodal Online Test-Time Adaptation for CLIP, ICCV 2025

Questions for authors

-Which model architecture is used? How is it pre-trained and how is it updated?

  • For visual tasks, Pretrained visual encoder such as DINOv2[1] , tends to capture better domain-invariant visual features. Is the performance degradation considerable also in these setting?

  • Can't we just apply online-fine tuning or domain reweighting with larger-sized models?

Also see weaknesses.

[1] DINOv2: Learning Robust Visual Features without Supervision

Reviewer_J7HoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an online domain adaptation algorithm that can handle partial domain observability and data imbalance. Experimental results shows clear improvements on Growning Circle, CompCars, and TPT-48 datasets.

Critical strengths
  1. This paper addressed a meaningful problem for online domain adaptation under partial observability and imbalanced data.
  2. The experimental results are promising. The ablation studies indicate the effectiveness of each components.
Critical weaknesses
  1. Although the proposed methods look better than VDI, this paper mainly compares against traditional domain adaptation methods and an online extension of VDI. More comparisons with online domain adaptation methods are missing.
  2. The paper is motivated by personalized ads and recommendation system. However, this paper didn't do any experiments on datasets related to personalized ads and recommendation.
  3. This paper also emphasizes dynamics and resource-limited environments, but computational efficiency is not quantitatively evaluated.
  4. The proposed method may not be robust to abrupt domain shift.
Questions for authors
  1. Could you compare with more online domain adaptation methods, such as "Online Multi-Source Domain Adaptation Through Gaussian Mixtures and Dataset Dictionary Learning" and "CONTRAST: Continual Multi-source Adaptation to Dynamic Distributions"?
  2. Could PIVOT be evaluated on a real-world ads dataset to validate its practical applicability?
  3. Could authors report adaptation time and memory usage?
  4. Although discussed in limitation, could authors evaluate robustness under abrupt distribution shifts and compare with other methods?

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

Related papers

Discussion (0)

Sign in to comment.