acceptodds
Withdrawn from ICLR 2027

Learning What to Transfer? Self-Supervised Disentanglement of Transferable Semantics for Source-Free Multivariate Time Series Domain Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We consider source-free domain adaptation (SFDA), which transfers a source-trained model to an unlabeled target domain without retaining source samples. For multivariate time series (MTS), target self-supervision can reveal temporal, spectral, and channel-relation structure, but consistency within the target domain does not necessarily imply relevance to the source task. Without source data for direct verification, indiscriminately transferring such structure may retain target-specific variation at the expense of task-relevant semantics. We propose self-supervised transferable semantic disentanglement (STSD) for source-free MTS domain adaptation. During source training, we decompose temporal, spectral, and channel-relation views into a shared semantic representation and view-specific components, and calibrate normalized classifier directions as persistent class anchors. We then use the frozen anchors to form a source semantic scatter operator and disagreement among unlabeled target views to form a target self-supervised variation operator. STSD solves a generalized spectral problem whose leading eigenspace preserves source class separation while suppressing target cross-view disagreement. The resulting eigenspace induces complementary transferable and target-local projectors, so semantic disentanglement follows from the geometry of the construction rather than from an auxiliary decoupling loss. Target prototypes are matched with the frozen source anchors only in the transferable subspace through entropic optimal transport. The target encoder is finally updated with a single semantic objective, while the spectral subspace and transport plan are refreshed from the current target representations. Extensive experiments on three standard MTS benchmarks demonstrate the effectiveness and superiority of STSD. The code is publicly available at https://github.com/Sky-byte-box/STSD_Code.

Primary area:
learning on time series and dynamical systems
Keywords:
Source-Free Domain Adaptation, Multivariate Time Series, Self-Supervised Learning, Transferable Semantics, Semantic Disentanglement

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_p3VyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes STSD for source-free multivariate time-series domain adaptation. During source training, the method learns shared representations from temporal, spectral, and channel-relation views and calibrates classifier weights as class anchors. During target adaptation, it combines source-anchor scatter with target cross-view disagreement through a generalized eigenvalue problem to identify a transferable subspace. Confidence-weighted target prototypes are then matched to projected source anchors using entropic optimal transport, and confidence-weighted pseudo-label training is finally performed. Experiments report improvements over existing methods on three standard MTS benchmarks.

Critical strengths
  1. The motivation is clear: consistency across target views does not necessarily indicate relevance to the source classification task.

  2. The proposed method is complete and resonable: Combining source-anchor discrimination and target cross-view disagreement in a generalized spectral criterion provides a principled formulation for subspace selection.

Critical weaknesses
  1. The novelty beyond existing source-semantic guidance is insufficiently established. The paper motivates its contribution through task relevance, but its task-related information is supplied by calibrated source classifier anchors. From my view, it is just like a kind of prototype-based methods[1,2]. Furthermore, STSD requires specialized multi-view source training and anchor calibration, limiting its demonstrated applicability to existing pretrained source models.

  2. The task-relevance claim is not convincing. Source-anchor discrimination and target cross-view consistency provide a reasonable inductive bias, but do not establish transferable task relevance: source anchors may encode spurious correlations, while view-specific information may remain task-relevant. Orthogonality guarantees only a geometric decomposition. The ablations and visualizations demonstrate utility rather than semantic separation, and Theorem 1 assumes favorable alignment without proving that STSD achieves it. Direct analysis of the retained and discarded subspaces is needed to support the claim.

  3. The class-mass term appears ineffective under the stated uniform source prior. In Eq. (9), if πjs=1/K\pi_j^s=1/K, the term ρ∣πˉkt−πjs∣\rho|\bar\pi_k^t-\pi_j^s| is constant across columns. Since the row marginals are fixed, its contribution to the transport objective is constant over all feasible plans. Consequently, changing ρ\rho cannot change the optimal transport solution. This conflicts with the claimed benefit in Fig. 3(d), unless the experiments use an exported nonuniform source prior or a different formulation.

[1] Source-free domain adaptation with class prototype discovery

[2] Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation

Questions for authors
  1. Can you report the selected subspace dimensions, projected-anchor margins, and transport-induced anchor deviations, and evaluate whether Theorem 1’s conditions hold in practice?

  2. What prevents the view-specific projectors from collapsing to zero? Can you report residual-feature variance or effective rank to demonstrate information retention?

Reviewer_t9JLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a source-free domain adaptation method for multivariate time series.

Critical strengths

The contributions seem to be relevant and are clearly presented. The paper is well written. Experiments, although not comprehensive, are convincing.

Critical weaknesses

The paper is missing some related methods in addressing source-free domain adaptation for time series, mainly:

  • Gaurav Patel, Chris Sandino, Behrooz Mahasseni, Ellen Zippi, Erdrin Azemi, Ali Moin, Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement, ICLR 2025
  • Yucheng Wang, Peiliang Gong, Min Wu, Felix Ott, Xiaoli Li, Lihua Xie, Zhenghua Chen. Temporal Source Recovery for Time-Series Source-Free Unsupervised Domain Adaptation. ArXiv 2409.19635v2

It would have been relevant to also include these papers in the experimentation analysis.

In the experiments, it is not clear how the hyperparameters were selected (besides being selected on source data). These include the choice of the fixed values of the cosine temperatures, disagreement temperature, class-mass coefficient, spectral ridge, eigenvalue-energy threshold, Sinkhorn regularization, and EMA coefficient.

Questions for authors

none

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

This paper proposes Self-Supervised Transferable Semantic Disentanglement (STSD) for source-free multivariate time-series domain adaptation. In the source pre-training phase, the framework constructs temporal, spectral, and channel-relation views, learns shared and view-specific representations, and calibrates classifier directions to serve as class anchors. At the target adaptation phase, it formulates a generalized eigenvalue problem that balances source-anchor separation with target cross-view disagreement. Experiments demonstrate superior macro/micro F1 scores across multiple benchmarks compared to baseline methods.

Critical strengths

Practical Problem Setting: The paper addresses source-free time series domain adaptation, a realistic and important setting where source data privacy and storage limitations preclude access to raw source samples during adaptation.

Critical weaknesses
  1. Undefined Parameters in Sensitivity Analysis: Section 4.4 and Figure 4 report parameter sensitivity for ξ\xi, gg, ε\varepsilon, and ρ\rho, but these parameters are nowhere clearly defined or contextualized within the main text/formulations. Without explicit mathematical definitions and physical meanings, it is impossible to evaluate the technical significance or practical implications of this sensitivity analysis.
  2. Gap Between Geometric Orthogonality and Semantic Disentanglement: The claim that geometric subspace separation corresponds to semantic disentanglement lacks rigorous theoretical and empirical justification. Enforcing orthogonality in Eq. (6) merely ensures linear independence between components; it does not guarantee that the selected subspace isolates transferable class semantics while the orthogonal complement captures target-specific variations. The source-discriminative, cross-view-consistent directions could still encode spurious domain factors, while the discarded subspace might retain crucial semantic signal.
  3. Inconclusive Visual Evidence in Figure 4: Decreased feature overlap in t-SNE visualizations (Figure 4) does not directly demonstrate the successful transfer of source semantics. To substantiate this claim, the authors should provide quantitative evaluations (e.g., probing class accuracy, domain discriminability, and predictive utility across both the selected subspace and its complement).
  4. Inconsistent Motivation for Optimal Transport (OT): The paper motivates the OT module by citing potential index permutations among target pseudo-labels. However, target probabilities are computed directly against source class anchors with fixed semantic identities, and target prototypes are constructed via these class-indexed probabilities. Because these prototypes inherit fixed class semantics, they do not suffer from the arbitrary label/index ambiguity typical of unsupervised clustering. While class assignment errors (misclassifications) certainly occur, this does not equate to arbitrary permutation, making the theoretical motivation for OT fundamentally misaligned with the method's design.
  5. Limited Benchmark Scope and Selection Bias: The empirical evaluation relies on only three datasets (two for Human Activity Recognition and one for Sleep Staging). More critically, the authors evaluate only two source-to-target subject pairs per dataset without explaining the selection criteria. Simply asserting that these tasks are “representative” is insufficient; a comprehensive evaluation across all subject pairs (including challenging/unfavorable pairs) and additional time-series domains is required to establish robustness against diverse domain shifts.
  6. Lack of Implementation Details and Reproducibility: Due to the absence of an appendix, critical implementation details, hyperparameter configurations, and architectural specifications are omitted, making it difficult to verify the results or assess the model's generalization capabilities.
  7. Besides, I did not see any AI Use Statement, Limitation discussions, or checklist in the pdf.
Questions for authors
  1. Parameter Definitions: What are the exact mathematical definitions and roles of the parameters ξ\xi, gg, ε\varepsilon, and ρ\rho shown in Figure 4? Why were these specific parameters chosen for sensitivity analysis over other hyperparameters?
  2. Subspace Validation: What empirical or theoretical evidence proves that the selected subspace contains transferable semantics while its orthogonal complement captures target-local variation? Could you provide a quantitative analysis (e.g., classification accuracy and domain discriminability) for both subspaces?
  3. Optimal Transport Rationale: Since target prototypes are constructed using probabilities tied to fixed source anchors, why would their index alignment undergo arbitrary permutations? What specific alignment failure does the OT module resolve that standard index matching cannot?
  4. Task Selection and Robustness: What were the exact criteria for selecting the six reported transfer tasks? Can you provide comprehensive results across all cross-subject pairs (including worst-case scenarios)?
  5. Implementation Details: Could you provide the complete experimental and implementation setup (e.g., neural architecture, optimization details, learning rates, and random seeds) to facilitate exact reproducibility?

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

Related papers

Discussion (0)

Sign in to comment.