acceptodds
Under review as a conference paper at ICLR 2027

Learning Style-Invariant Temporal Semantics for Source-Free Time-Series Domain Adaptation

Abstract

We study source-free time-series domain adaptation (SFTSDA), where a source-trained predictor is transferred to an unlabeled target domain while source samples are unavailable. Existing approaches improve transfer through imputation, time–frequency modeling, and target self-training, yet target-specific temporal style may remain entangled with class-relevant semantics, causing style variation to be interpreted as class evidence and subsequently reinforced through pseudo-label adaptation. We propose style-invariant temporal semantics (SITS), which formulates semantic invariance through target-style orbit learning. Target-only channel and spectral statistics define closed-form style transformations, while evolving temporal semantic neighborhoods determine which transformations are admissible for each sample, forming a target-style orbit. SITS then learns relation-aware temporal semantics by maximizing the worst-case class margin over each orbit. Progressive pseudo-posterior refinement updates the semantic neighborhoods used to construct the next orbit. We further show that a positive orbit margin preserves the predicted class over the covered target-style family, providing a direct link between the orbit-wise objective and style-invariant temporal semantics. Experiments on three real-world time-series datasets evaluate the proposed SITS framework against representative source-free and test-time adaptation baselines. The code is available at https://github.com/Sky-byte-box/SITS_code.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Source-Free Domain Adaptation, Time-Series Classification, Style-Invariant Representation Learning, Temporal Semantic Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_a8T4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies source-free time-series domain adaptation and argues that target-specific style (amplitude, channel offset, spectral emphasis) can be mistaken for class evidence and then reinforced by pseudo-label self-training. The authors compare their method against multiple baseline methods on three data sets.

Critical strengths
  • The paper is well written with clear problem statement and method presentation

  • The proposed method is simple to implement.

Critical weaknesses
  • Theorem 1 follows directly from Lipschitz continuity and a margin argument. So, I wonder whether it's overly strong to be called"theorem".
Questions for authors

Please consider weakening the Theorem 1.

3 comments
AuthorsOpenReview ↗

Response to Reviewer a8T4

Thank you so much for your positive and encouraging feedback. Below, we address your comments and suggestions in details:

W1 & Q1: Scope and designation of Theorem 1.

A. We appreciate the reviewer's observation. Theorem 1 establishes a sufficient condition for orbit-wise predicted-class preservation based on Lipschitz continuity and a positive class margin. To reflect its precise theoretical scope, we have relabeled it as Proposition 1, while retaining the original statement and proof. Specifically, when the target-style orbit forms an ϵi\epsilon_i-cover of the plausible target-style family in semantic space and the minimum orbit margin exceeds 2Lcϵi2L_c\epsilon_i, the predicted class is preserved throughout that family. Proposition 1 explicitly connects the orbit-wise margin objective to this conditional preservation guarantee.

Thanks for the response. I'll continue supporting this paper.

We sincerely thank the reviewer for the positive feedback and continued support.

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

This paper focuses on Source-Free Time-Series Domain Adaptation (SFTSDA), which transfers a pre-trained source-domain model to an unlabeled target domain without access to source data. The authors argue that target-domain-specific temporal styles are entangled with category-related semantic features. This causes the model to misinterpret style variations as categorical evidence during pseudo-label adaptation, and such erroneous predictions are continuously reinforced within the self-training loop based on pseudo labels. To tackle this issue, the paper proposes a novel method named SITS (Style-Invariant Temporal Semantics). Its core mechanisms include:

  1. Defining closed-form style transformations using channel and spectral statistics of the target domain, and constructing target style orbits to delineate semantic neighborhoods;
  2. Maximizing the worst-case category margin on the orbits to learn relation-aware temporal semantics;
  3. Performing progressive pseudo-posterior refinement by combining source prototypes and semantic graph propagation. In addition, the authors theoretically prove that a positive orbit margin preserves the stability of predicted categories over the covered family of target styles. Extensive evaluations are conducted across six cross-domain setups on three real-world time-series datasets: UCIHAR, MFD and Sleep-EDF. Experimental results show that SITS achieves the best Macro-F1 score on all setups and outperforms existing competing methods.
Critical strengths
  1. Practically meaningful problem formulation. Source-free time-series domain adaptation is highly demanded in industrial fault diagnosis, healthcare monitoring and other real-world scenarios. However, source data is often unavailable due to privacy and data compliance constraints. This work identifies a critical failure mode in SFTSDA: style variations are misdiagnosed as semantic evidence and amplified during pseudo-label self-training.
  2. Reasonable innovation and targeted method design. The proposed concept of target style orbit is elegant. Closed-form transformations are built upon inherent channel and spectral statistics of time-series signals, eliminating the need to train extra style generators or complex disentanglement branches, which improves training efficiency and stability. A dynamic semantic affinity graph is introduced to constrain the range of style exchange and avoid blind style injection across unrelated categories.
  3. Solid theoretical guarantees. Unlike many domain adaptation works that only rely on empirical heuristics, this paper provides theoretical support for SITS. It proves that under the positive orbit margin condition, the model maintains consistency of category predictions across the family of target styles. This strengthens the credibility of the proposed approach and offers valuable theoretical insights for future research.
  4. Comprehensive experimental validation. Evaluations are performed on three distinct time-series datasets covering human activity recognition, machine condition monitoring and physiological signals, verifying the generalization capability of SITS. Thorough comparisons against 8 baselines including TENT, EATA, SAR, RoTTA, CATI and LEMON are provided.
Critical weaknesses
  1. Limited performance gains without statistical significance tests. Although SITS achieves the best results on all setups, the average Macro-F1 improvement is only 0.5–0.7 percentage points compared with the strongest baseline CATI. Such marginal gains may fall within random fluctuations for time-series classification tasks.
  2. Numerous hyperparameters and insufficient discussion of computational overhead. SITS introduces multiple hyperparameters including Knn, ξ, τᵤ, τₚ, γ, τₘ, which substantially increase tuning difficulty in practical deployment. Moreover, constructing style orbits, semantic graph propagation and progressive pseudo-posterior refinement incur additional computation steps. The paper lacks comparisons of training/inference time and GPU memory consumption against baseline methods. Computational efficiency is a critical concern for edge-based time-series applications with tight resource budgets.
  3. Coupled variables in ablation studies. The ablation setting “w/o target-style orbit” removes multiple components simultaneously, including target style orbit construction and relation-aware margin maximization. This makes it impossible to precisely attribute performance degradation to either the absence of style orbits or the removal of margin loss. Finer-grained ablations that decouple these components are recommended.
Questions for authors
  1. For certain time-series tasks (e.g., EEG signals in Sleep-EDF), frequency-domain energy shifts or amplitude drifts may inherently correlate strongly with semantics of pathology or sleep stages. When distribution shifts between source and target domains exceed the coverage of closed-form transformations, or when certain frequency variations carry categorical semantics, would SITS’s core assumption that “style does not alter category” introduce excessive smoothing or semantic ambiguity?
  2. The framework involves many hyperparameters (e.g., Knn, ξ, τᵤ, τₚ, γ, τₘ). While the main text presents parameter sensitivity analysis on Sleep-EDF (Fig.3), do these hyperparameters require extensive re-tuning across different datasets?
  3. Could you report the mean and standard deviation of SITS and CATI over multiple independent runs (e.g., 5 or 10 seeds) across all six cross-domain setups? Have you conducted statistical significance tests to verify that the 0.5–0.7% improvement is not merely a random artifact?
4 comments
AuthorsOpenReview ↗

Response to Reviewer yNrV—Part I

We are grateful for the reviewer’s constructive remarks and suggestions. We respond to each point in the following.

W1 & Q3: Statistical reliability of performance improvements.

A. Table 1 reports mean ±\pm SD over three runs. To address the reviewer's request, we additionally evaluated SITS and CATI over ten independent runs on all six original transfers. Appendix Table 6, reproduced below, provides mean ±\pm SD, mean MF1 differences with unadjusted 95%95\% confidence intervals, and Holm-adjusted pp-values for six comparisons. SITS has higher mean MF1 on all six transfers (gains of 0.280.28–1.191.19 percentage points), while all Holm-adjusted pp-values exceed 0.050.05 (0.0620.062–0.6550.655). Thus, we report observed mean gains, not statistically established superiority on individual transfers. Appendix Tables 3–5 additionally extend evaluation to 27 transfers, with SITS obtaining the highest mean MF1 on 23, providing complementary evidence across tasks.

Table 6. Ten-run Macro-F1 (%) statistics for SITS and CATI.

TransferSITS (10 runs)CATI (10 runs)ΔMF1\Delta\mathrm{MF1} (95%95\% CI)Adjusted pp
UCIHAR 9$\rightarrow$1884.43$\pm$2.6684.15$\pm$2.68+0.28 [-1.09,1.65]0.655
UCIHAR 12$\rightarrow$1667.99$\pm$3.5066.80$\pm$3.59+1.19 [-0.45,2.83]0.405
MFD 1$\rightarrow$291.48$\pm$0.7790.98$\pm$0.55+0.50 [0.15,0.85]0.062
MFD 1$\rightarrow$089.95$\pm$1.7589.35$\pm$2.09+0.60 [-0.46,1.66]0.465
Sleep-EDF 12$\rightarrow$565.60$\pm$1.2864.87$\pm$1.16+0.73 [0.05,1.41]0.190
Sleep-EDF 9$\rightarrow$1471.38$\pm$1.2370.90$\pm$1.16+0.48 [-0.06,1.02]0.301

Note. ΔMF1\mathrm{MF1} = MF1\mathrm{MF1}(SITS) − MF1\mathrm{MF1}(CATI). The 95% CIs are unadjusted. The reported pp-values are Holm-adjusted across six comparisons.

W2 & Q2: Hyperparameter selection and computational cost.

A. The six parameters KnnK_{\mathrm{nn}}, ξ\xi, τu\tau_u, τp\tau_p, γ\gamma, and τm\tau_m govern neighborhoods, refinement, propagation, and orbit optimization. Their values were selected using the Sleep-EDF sensitivity analysis (Fig. 3) and applied to other transfers without task-specific re-tuning. Baselines use the hyperparameters recommended in their original publications. Target adaptation remains unlabeled. We added complexity analysis after Algorithm 1 (Eq. 12), including exact neighbor search (O(N^2d)), graph diffusion (O(T_gNkK)), orbit generation (O(NkCL\log L)), and multi-view encoding (O(N(k+3)F_{\mathrm{enc}})) per round. Table 7 below reports measured costs for SITS and all 11 baselines on MFD 1→21\rightarrow2.

Table 7. Measured computational costs of SITS and all 11 baselines on MFD 1→21 \rightarrow 2.

MethodSource training (min)Target adaptation (min)Inference (s/1k)Peak GPU memory (GiB)
Source Only8.4N/A0.401.5
TENT9.84.80.412.2
SHOT-IM10.49.00.443.2
EATA9.74.60.412.8
SAR10.58.80.453.3
RoTTA9.911.70.523.4
ACCUP10.316.70.534.1
CATI19.414.90.525.1
LEMON18.417.20.745.2
DA-WS15.216.80.463.5
TransPL27.821.81.086.2
SITS10.627.50.516.7

Compared with CATI, SITS has higher adaptation time (27.5 vs. 14.9 min) and peak GPU memory (6.7 vs. 5.1 GiB), but comparable inference latency (0.51 vs. 0.52 s/1k). This reflects additional orbit and semantic-graph operations; GPU results alone do not establish edge-device efficiency.

AuthorsOpenReview ↗

Response to Reviewer yNrV—Part II (Continued)

W3: Disentangling the effects of orbit construction and margin optimization.

A. Table 2 adds Identity-orbit (identical views, retained margin; single-view equivalent) and Worst-case orbit cross-entropy (retained orbit and worst-case aggregation, cross-entropy replacing margin). Results follow.

Table 2. Mechanism-focused ablations (MF1 %, mean ±\pm SD over three runs).

VariantUCIHARMFDSleep-EDF
w/o target-style orbit68.65$\pm$3.1083.90$\pm$1.5559.80$\pm$1.72
Identity-orbit69.67$\pm$3.0884.77$\pm$1.7360.74$\pm$1.57
Worst-case orbit cross-entropy73.51$\pm$2.5188.35$\pm$1.2265.71$\pm$1.13
Mean-orbit objective72.85$\pm$2.6587.75$\pm$1.1464.70$\pm$1.10
Full SITS75.95$\pm$2.3690.65$\pm$0.9268.40$\pm$0.84

Full SITS exceeds Identity-orbit by 6.28/5.88/7.66 points and worst-case orbit cross-entropy by 2.44/2.30/2.69 on UCIHAR/MFD/Sleep-EDF. Mean-orbit objective isolates aggregation. The original w/o orbit removes both mechanisms; its drop is a joint effect.

Q1: Semantic preservation under category-correlated style variation.

A. SITS does not indiscriminately treat amplitude or frequency changes as category-preserving styles. Eq. (5) restricts style donors through mutual semantic proximity and posterior agreement, reducing the risk of exchanging class-incompatible spectral patterns. Eqs. (6)–(7) further preserve discriminative class evidence by maximizing the weakest margin over these admissible target-style orbits, rather than enforcing unconditional smoothing. Proposition 1 establishes predicted-class preservation over a covered target-style family when the orbit margin exceeds 2Lcϵi2L_c\epsilon_i. Its guarantee is conditional on semantic coverage and a positive margin, rather than assuming invariance to arbitrary frequency changes or distribution shifts. Thus, SITS targets semantically compatible style variations while retaining class-discriminative temporal information.

The authors' responses provide reasonably solid support for the paper's methodology, and I will continue to support this paper.

We sincerely thank the reviewer for the positive feedback and continued support.

Reviewer_R2UTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I tried to understand the algorithm. Please feel free to correct me if my understanding is wrong. This paper proposes an algorithm (SITS) for source-free time-series domain adaptation (SFTDA). SFTDA works by updating the source encoder. The source classifier is kept frozen, and only the encoder-side parameters are updated. The encoder and the classifier are used to make predictions on the unlabeled target, without access to the source samples at all times. SITS works as follows. (1) All target samples are encoded with the source encoder using a self-attention-like mechanism (Eq1), leading to a representation cic_i. Eq 1 is used to emphasize the relative difference between tokens, emphasizing the local transition. (2) It then constructs a target class prototype from the unlabeled target samples. Using this prototype and the initial class prediction, the pseudo labels are constructed and refined through an iterative process. The class predictions (from the source model and the prototypes) are softly mixed and smoothed through label propagation. To mitigate the effect of misclassified class predictions, highly confident samples are selected for label propagation. The mixing ratio (threshold to select confident samples) is increased through several rounds. (3) For each sample ii, select a k-nearest neighbor that also has high class prediction similarity. These neighbors are used to construct a target-style orbit (basically, augmentations of sample ii that maintain the content but with a different style). The orbit and the samples are used to update the source encoder. This is repeated for several rounds.

Critical strengths
  • The intuition to make an orbit (i.e., augmentations) of a target sample using style invariant mechanism (similar to adaIN) is novel and interesting. Best to my knowledge, I haven’t seen other time-series works that utilize similar approach.

  • The paper was quite hard to follow with the equations and notations which were dense. However, with some effort, I would say that it is understandable and the intuition behind the algorithm is novel and interesting. It would be of help for researchers trying to construct a representation, augmentation for unlabeled time series samples. I also think it is a strong fit for ICLR which is about learning representations.

Critical weaknesses
  • The paper does not specify how the hyperparameters were selected. Was a small subset of target data used for tuning? Also, were the hyperparameters of the baselines tuned?

  • If the initialization of pseudo labels for the target were wrong at the beginning, the samples under the incorrectly labeled class are likely to share similar styles and they might end up updating the source encoder in the wrong direction.

  • The notations and equations are extremely dense, and it is very hard to follow. For instance, cic_i seems to be the core representation, but its definition is in eq 3. As a reader, it would be more easier to follow if this was defined somewhere close to eq1. I had to go back and forth several times to find where the notation were placed.

  • The computation required seems rather complex. The computation and algorithm complexity need to be explained.

  • Why did the paper only test on transfers of two groups for each dataset? Previous literature usually tests on 10 different source-target pairs. The results on two groups may not generalize.

  • Some missing DA baselines that utilize pseudo-labels.

[1] Multi-Source Deep Domain Adaptation with Weak Supervision for Time-Series Sensor Data

[2] TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation

Questions for authors

(Minor) white space is needed in Fig 2 caption.

The relation between channels, sequence length, and temporal tokens should be described more explicitly. Is T = C x L? or are the channels aggregated? If the channels are aggregated, is the transition considered in a multi-channel manner? These descriptions are missing. Also, ziz_i seems to be the latent embeddings after attention (eq. 1). The notation is not defined.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R2UT—Part I

Thank you for your insightful comments and for taking the time to review our paper. We respond to your suggestions point by point as follows.

W1: Hyperparameter selection and baseline tuning.

A. We chose the SITS hyperparameters (KnnK_{\mathrm{nn}}, ξ\xi, τu\tau_u, τp\tau_p, γ\gamma, and τm\tau_m) with reference to the sensitivity analysis on Sleep-EDF in Fig. 3 and share these settings on the other source-to-target transfers. Thus, we did not separately tune the shared parameters using a labeled target subset for each transfer. Target adaptation itself uses only unlabeled target data. For all baselines, we adopted the hyperparameter settings recommended in their original publications.

W2: Incorrect initial pseudo-labels.

A. SITS does not directly update the encoder using all initial pseudo-labels. Importantly, style similarity alone does not determine semantic neighborhoods. Eq. (5) requires mutual nearest-neighbor consistency in semantic space and posterior agreement, with confidence-based weighting. Eqs. (8)–(9) progressively refine the initial predictions through reliability-weighted prototypes and semantic graph propagation. Moreover, class-balanced confidence selection restricts the samples used for the encoder update in Eq. (10). These mechanisms limit the direct influence of unreliable initial predictions on subsequent adaptation. Table 2 further shows that removing progressive semantic refinement decreases mean MF1 by 6.50, 6.10, and 7.70 points on UCIHAR, MFD, and Sleep-EDF, respectively.

W3 & Q1 & Q2: Clarification of notation, temporal tokens, and Figure 2 caption.

A. For xi∈RC×Lx_i\in\mathbb{R}^{C\times L}, CC and LL denote the channel count and sequence length, respectively. The 1D convolutional encoder integrates information across channels and produces Hi∈RT×dhH_i\in\mathbb{R}^{T\times d_h}, where TT is the number of encoded temporal positions, rather than C×LC\times L. Thus, the adjacent-token differences in Eq. (1) capture temporal transitions in the channel-integrated feature space. Eq. (1) yields zi∈Rdhz_i\in\mathbb{R}^{d_h}, the relation-aware sequence embedding, which is projected into the classifier-facing semantic code ci=norm⁡(Wczi)c_i=\operatorname{norm}(W_c z_i). We have clarified these definitions around Eq. (1) in Section 3.1 and corrected the spacing in the Figure 2 caption.

W4: Computational complexity.

A. We have added the computational complexity analysis after Algorithm 1. Let NN, dd, kk, KK, CC, LL, and FencF_{\mathrm{enc}} denote the number of target samples, semantic dimension, neighbor cap, number of classes, channels, sequence length, and per-sequence encoding cost, respectively. In each adaptation round, exact neighbor search, graph diffusion, FFT-based style transfer, and encoding the original, two mild, and up to kk orbit views require O(N2d)O(N^2d), O(TgNkK)O(T_gNkK), O(NkCLlog⁡L)O(NkCL\log L), and O(N(k+3)Fenc)O(N(k+3)F_{\mathrm{enc}}), respectively. Including one-time style statistics and classifier evaluation, the overall complexity is O(NCLlog⁡L+R[N2d+TgNkK+NkCLlog⁡L+N(k+3)Fenc+N(k+1)Kd])O(NCL\log L + R[N^2d + T_gNkK + NkCL\log L + N(k+3)F_{\mathrm{enc}} + N(k+1)Kd]). With k=8k=8, Tg=10T_g=10, and R=5R=5, each orbit contains at most nine sequences, while exact neighbor search remains the main quadratic-in-NN component.

In addition, Appendix C, Table 7 further reports measured computational costs for SITS and all 11 baselines on MFD 1→21\rightarrow2.

Table 7. Measured computational costs of SITS and all 11 baselines on MFD 1→21 \rightarrow 2.

MethodSource training (min)Target adaptation (min)Inference (s/1k)Peak GPU memory (GiB)
Source Only8.4N/A0.401.5
TENT9.84.80.412.2
SHOT-IM10.49.00.443.2
EATA9.74.60.412.8
SAR10.58.80.453.3
RoTTA9.911.70.523.4
ACCUP10.316.70.534.1
CATI19.414.90.525.1
LEMON18.417.20.745.2
DA-WS*15.2*16.8*0.463.5
TransPL27.821.81.086.2
SITS10.627.50.516.7

SITS incurs higher target adaptation time and peak GPU memory than the compared methods on MFD 1→21\rightarrow2, reflecting additional target-style orbit construction and semantic graph operations. Nevertheless, its inference latency remains comparable to most baselines, indicating that the main computational overhead occurs during target adaptation rather than inference.

AuthorsOpenReview ↗

Response to Reviewer R2UT—Part II (Continued)

W5: Generalizability across transfer directions.

A. We expanded the evaluation from two to nine source-to-target transfers per dataset, adding seven non-overlapping transfer tasks for each of UCIHAR, MFD, and Sleep-EDF. These additional tasks, summarized in Appendix Tables 3–5, were selected from transfer directions reported in prior studies. SITS achieves the highest mean MF1 on 17 of the 21 additional transfers. On the remaining four transfers (UCIHAR 6→23 and 18→21, MFD 2→1, and Sleep-EDF 10→7), its mean MF1 is within 0.13 percentage points of the best-performing method. Together with its highest MF1 on the six original transfers, SITS achieves the best mean MF1 on 23 of the 27 evaluated transfers.

W6: Missing pseudo-labeling baselines.

A. We have additionally included the reviewer-suggested DA-WS and TransPL methods in Tables 1 and 3–5, covering 27 source-to-target transfers across UCIHAR, MFD, and Sleep-EDF. SITS achieves higher mean MF1 than both methods on all 27 transfers.

Table R1. Summary of transfer coverage and pseudo-labeling baseline comparisons (from Tables 1 and 3–5).

DatasetSITS best (7 added)SITS > DA-WS (all 9)SITS > TransPL (all 9)
UCIHAR5/79/99/9
MFD6/79/99/9
Sleep-EDF6/79/99/9
Total17/2127/2727/27

Thank you for the thorough reply and explanation. I re-read the revised PDF, and I believe that many of my concerns have been addressed. I have adjusted my score accordingly. However, I still think that the paper itself is quite dense with notations and equations, making it slightly hard to follow. I would strongly advise the authors to move some of the main equations to the Appendix to make the main text more readable.

We thank the reviewer for the positive reassessment and constructive suggestion. We appreciate the concern regarding the density of mathematical notation. We will improve the readability of the main text by retaining the essential methodological equations and moving selected supplementary formulations and derivations to the Appendix, while preserving the technical content.

Comments on the paper

1 comment
Area_Chair_1RZoOpenReview ↗

Reminder: Please Participate Actively in the Discussion

Dear Reviewers,

Please actively participate in the discussion by responding to the authors’ replies and engaging with your fellow reviewers. This discussion is essential for clarifying concerns, resolving differences, and reaching a well-supported consensus on the final recommendation.

Thank you for your time and valuable contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.