acceptodds
Withdrawn from ICLR 2027

Adaptive Spectral-Koopman Dynamics Modeling for Temporal Domain Generalization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal Domain Generalization (TDG) has emerged to address real-world streaming data with distribution shifts over time. However, existing methods are either prone to overfitting to domain-specific noise in the data space or become overly complex and less interpretable in the parameter space. To bridge these gaps, we propose AdaSpecK, a spectral-Koopman framework with adaptive context extraction for TDG. To mitigate noise fitting to irregularly sampled domains, we introduce spectral-regularized Koopman dynamics modeling, which applies spectral-aware filtering in the latent space to extract denoised low-frequency trajectories and learn a Koopman operator to model the system dynamics in a linearized space. To model complex historical environments under non-stationarity, we design a context-informed heterogeneous pattern extraction mechanism. Specifically, we employ a target-conditioned attention module to attend to distinct past windows, producing a dynamic, target-specific historical summary. By constructing an environmental signature from the current evolutionary pattern, our model adaptively perceives which aspects of the past context are most informative for future prediction via a learned router. Extensive experiments on eight diverse classification and regression benchmarks demonstrate that AdaSpecK achieves state-of-the-art performance. The code and datasets are available at https://anonymous.4open.science/r/Ada-Spec-K.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
temporal domain generalization

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_HX52OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AdaSpecK for temporal domain generalization. It applies spectral filtering to domain-level feature trajectories, learns their evolution through a Koopman operator, and combines target-conditioned historical attention with environment-guided expert routing. Experiments on eight classification and regression benchmarks report improvements over existing TDG methods.

Critical strengths
  1. The method combines environment evolution with sample-level prediction in a coherent overall framework. Modeling compact environment representations is a reasonable alternative to forecasting full model parameters.

  2. Evaluation spans eight benchmarks and includes repeated runs and ablations.

Critical weaknesses
  1. The novelty is limited. FreKoo (Yu et al., 2025) already addresses long-term evolution and domain-specific uncertainties through spectral decomposition and low-frequency Koopman modeling. AdaSpecK’s filtering, encoding–transition–decoding pipeline, and transition and reconstruction objectives substantially overlap with this design. The paper does not sufficiently establish a distinct contribution beyond FreKoo, and the incremental additions fall short of the novelty expected at ICLR.

  2. Ordinary DFT over domain indices and a fixed transition zt+1=Kenvztz_{t+1}=K_{\mathrm{env}}z_t do not explicitly account for unequal time intervals. Moreover, Eq. (10) appears to produce a single context embedding; if this supplies only one key/value to Eq. (12), attention becomes independent of the target query. The amplitude and phase in Eq. (13) are also undefined for the real-valued, inverse-transformed representation. These ambiguities prevent verification of the claimed irregular-time modeling and target-specific context extraction.

  3. The ablations remove individual components but lack capacity-matched alternatives, such as a standard attention module or a single MLP replacing the experts. Performance across prediction horizons and controlled irregular-sampling settings would also help substantiate central claims.

  4. The transition and consistency losses constrain observed trajectories but do not guarantee stable future rollouts. The paper provides neither theoretical conditions nor empirical analysis of the Koopman operator’s spectrum and error growth over longer horizons.

Questions for authors
  1. What methodological contribution distinguishes the spectral–Koopman component from FreKoo beyond changing the modeled trajectory? Can controlled comparisons isolate the gains from environment-space modeling, historical attention, and spectral routing?

  2. Where do the actual time intervals enter the spectral transform and Koopman transition? How does the RK4 solver mentioned in the appendix relate to Eq. (6)?

  3. What is the shape of EctxE_{\mathrm{ctx}}, and how many historical keys/values are used in Eq. (12)? Please also specify how amplitude and phase are computed and confirm that the environmental signature uses only historical information and predicted future states.

Reviewer_E9m7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AdaSpecK, a temporal domain generalization framework combining spectral filtering, Koopman-based latent dynamics modeling, and adaptive historical context extraction. The method applies frequency-domain filtering to domain-level representations, learns linearized temporal dynamics through a Koopman operator, and integrates a target-conditioned attention mechanism with expert routing to construct historical context representations for future predictions.

Experiments on eight classification and regression benchmarks demonstrate improvements over existing TDG methods. Ablation studies examine the effects of the individual architectural components and regularization objectives.

Critical strengths

A meaningful combination of temporal modeling mechanisms. The integration of spectral filtering, latent dynamics, and historical context extraction provides a potentially useful framework for learning temporal representations under distribution shifts.

Broad empirical evaluation. The experiments cover eight benchmarks involving classification and regression. The method reports consistently competitive results, and the component ablations provide preliminary evidence that the proposed modules contribute to performance.

Critical weaknesses

The target-conditioned attention mechanism appears mathematically degenerate as formulated. Equation (10) aggregates the historical domain representations into a unified contextual embedding (EctxE_{\mathrm{ctx}}) through concatenation and an MLP. Equation (12) then performs cross-attention using projections of this contextual embedding as the key and value.

If (EctxE_{\mathrm{ctx}}) is a single embedding, as the formulation suggests, the attention mechanism contains only one key-value element. In this case, the softmax weight is always one, and the resulting attention output cannot depend on the target-conditioned query. Therefore, the mechanism would be unable to selectively attend to different historical windows, contrary to one of the paper's central claims.

This issue directly concerns the validity of the proposed context-informed heterogeneous pattern extraction mechanism. A precise tensor-level explanation is necessary to determine whether the problem is limited to the mathematical presentation or also affects the actual implementation.

The proposed spectral and Koopman formulations do not adequately address irregular temporal sampling, despite this being a central motivation. Equations (2)–(5) apply a conventional discrete Fourier transform along the sequence of observed domains. However, the standard DFT implicitly assumes uniform sampling intervals, while the paper explicitly motivates its method using irregularly sampled temporal domains.

Similarly, Equation (6) uses a fixed Koopman operator for transitions between consecutive domains without explicitly incorporating the elapsed time between observations. Consequently, transitions over different temporal intervals appear to be treated identically.

The paper does not establish how these operations correctly represent irregular temporal dynamics. Furthermore, the experiments do not directly evaluate robustness under systematically varied timestamp irregularity, making it difficult to assess whether the proposed method actually addresses its stated challenge.

Questions for authors

Clarification of the attention mechanism: What are the exact tensor dimensions of (EctxE_{\mathrm{ctx}}), the projected keys and values, and the attention weights in Equations (10)–(12)? If (EctxE_{\mathrm{ctx}}) contains only one contextual token, how can the attention operation perform target-dependent selection among historical windows? If multiple tokens are retained, please clarify how they are constructed.

Validation of irregular-time modeling: How does the proposed DFT-based filtering account for unequal temporal intervals? How does the Koopman transition incorporate the actual time difference between domains? Can the authors provide a controlled evaluation with varying degrees of timestamp irregularity to substantiate the claimed robustness?

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

This paper proposes AdaSpecK, a spectral-Koopman framework with adaptive context extraction for temporal domain generalization. It introduces spectral-regularized Koopman dynamics modeling to extract denoised low-frequency trajectories in a latent space and model their evolution through a learnable Koopman operator. To capture dynamic temporal patterns under non-stationarity, the framework employs target-conditioned attention to construct a target-specific historical summary. An environmental signature guides a learned router to selectively integrate historical information through multiple experts. Experiments on eight classification and regression benchmarks demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. The proposed method combining spectral filtering, Koopman dynamics, and adaptive historical context provides an interesting approach to temporal domain generalization.

  2. The empirical evaluation is comprehensive, covering eight classification and regression benchmarks with strong overall results. Ablation studies, sensitivity analyses, and decision-boundary visualizations provide further insight into the method.

  3. The paper presents its motivation and main components in a clear, well-structured manner.

Critical weaknesses
  1. The released implementation may differ from the paper’s description. Section 4.1 states that the DFT is applied along the temporal dimension, but _split_low_high applies it along the feature dimension of a tensor shaped as [batch, time, features]. In addition, _build_env_signature averages over features and reshapes the result so that the subsequent FFT operates on a length-one dimension, which cannot extract the temporal frequency and phase structure described in Section 4.2. Could the authors clarify these differences, or please point them out if I have misunderstood?

  2. The method is designed to let different future steps attend to different historical windows; however, main.py uses only the last training domain as context when predicting multiple future domains. With a single historical key/value, the attention weight is always 1, so the resulting historical summary is independent of the target query. Please point it out if I have misunderstood.

  3. Section 3 introduces arbitrary timestamps, but the methodology does not clearly distinguish domain indices from actual observation times.

  4. Eq. (6) describes a fixed linear transition, whereas the released code implements �0��0�, using time gaps computed within each input segment.

Questions for authors
  1. Could the authors clarify how the FFT axes and spectral-signature computation were computed?

  2. Could the authors clarify the temporal indexing and provide the complete transition equation, including how actual time gaps enter the model?

  3. Could the authors provide results and decision-boundary visualizations on an irregularly sampled rotating 2-Moons benchmark to assess whether the model captures the correct temporal dynamics?

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

Related papers

Discussion (0)

Sign in to comment.