cs.LGAug 23, 2026

FreKoo++: Learning Continuous Spectral Dynamics for Temporal Domain Generalization

Authors: En Yu, Xiaoyu Yang, Wei Duan, Guangquan Zhang, Jie Lu

Organizations: Australian Artificial Intelligence Institute (AAII), Faculty of Engineering and Information Technology, University of Technology Sydney (UTS), Australia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Temporal Domain Generalization (TDG) aims to learn from historical domains and generalize to unseen future distributions under concept drift. Nevertheless, prevailing TDG methods struggle with complex real-world streaming scenarios involving both multi-scale drift patterns (e.g., long-term periodicity intertwined with short-term incremental changes) and local uncertainties, especially in continuous settings where observations arrive irregularly. To address this limitation, we propose FreKoo++, a novel continuous spectral-dynamical framework that pioneers the unification of continuous Koopman modal dynamics with adaptive spectral disentanglement. Specifically, FreKoo++ maps source-domain parameters into a compact latent space, modeling their evolution as a superposition of learnable continuous modes where complex eigenvalues jointly encode oscillatory frequency and temporal growth or decay. This formulation naturally accommodates irregular timestamps and supports arbitrary horizon extrapolation without rigid discrete stepping. Furthermore, we propose a new adaptive soft spectral weighting mechanism backed by stability and spectral regularization, which automatically isolates persistent dominant dynamics from transient noise without relying on manual frequency thresholds. We derive modal approximation and generalization bounds that characterize how amplitude and eigenvalue estimation errors propagate with the prediction horizon. Extensive experiments on both discrete and continuous TDG benchmarks demonstrate that FreKoo++ achieves state-of-the-art performance under complex multi-scale drifts and irregular sampling.

Figures & tables

Fig. 1 : Illustration of the three TDG challenges on a 2-Moons stream. The underlying distribution exhibits multi-scale periodicity, while empirical parameter estimates are corrupted by domain-specific uncertainties. Additionally, the irregularly sampled observations violate the time-invariant discrete transition assumed by standard sequence models.
Fig. 2 : FreKoo framework. It decomposes model parameter trajectories into low- and high-frequency components via the Fourier transform. Koopman dynamics extrapolate the low-frequency component, while temporal-difference regularization suppresses unstable high-frequency variations.
Fig. 3 : FreKoo++ framework. (1) Training Phase: Model parameters from irregular timestamps are encoded into a continuous latent trajectory and fitted by Koopman modal dynamics. Adaptive soft weighting disentangles persistent dominant dynamics from transient noise, yielding an optimized continuous spectral representation via joint optimization. (2) Inference Phase: For an arbitrary target timestamp ts>tTt_{s}>t_{T} , FreKoo++ analytically extrapolates the persistent dominant component to tst_{s} , freezes the transient residual at tTt_{T} , and decodes their superposition to obtain target parameters θ^(ts)\hat{\theta}(t_{s}) .
MethodsClassificationRegression
2-Moons-C ↓\downarrowRot-MNIST-C ↓\downarrowTwitter ↑\uparrowYearbook ↓\downarrowCyclone ↓\downarrowHouse-C ↓\downarrow
Offline13.5 ±\pm 0.36.6 ±\pm 0.20.54 ±\pm 0.098.6 ±\pm 1.018.7 ±\pm 1.419.9 ±\pm 0.1
LastDomain55.7 ±\pm 0.574.2 ±\pm 0.90.54 ±\pm 0.1211.3 ±\pm 1.322.3 ±\pm 0.720.6 ±\pm 0.7
IncFinetune51.9 ±\pm 0.757.1 ±\pm 1.40.52 ±\pm 0.0111.0 ±\pm 0.819.9 ±\pm 0.720.6 ±\pm 0.2
IRM [ 1 ]15.6 ±\pm 0.28.6 ±\pm 0.40.53 ±\pm 0.118.3 ±\pm 0.518.0 ±\pm 0.819.8 ±\pm 0.2
V-REx [ 21 ]12.8 ±\pm 0.28.6 ±\pm 0.30.58 ±\pm 0.058.9 ±\pm 0.517.7 ±\pm 0.520.2 ±\pm 0.1
TABLE I : Performance comparison on CTDG datasets . The classification tasks report error rates (%) except for the AUC for the Twitter dataset. The regression tasks report MAE. ’-’ implies that the method does not support the task.
MethodsClassificationRegression
2-Moons-D ↓\downarrowRot-MNIST-D ↓\downarrowONP ↓\downarrowShuttle ↓\downarrowElec2 ↓\downarrowHouse-D ↓\downarrowAppliance ↓\downarrow
Offline22.4 ±\pm 4.618.6 ±\pm 4.033.8 ±\pm 0.60.77 ±\pm 0.1023.0 ±\pm 3.111.0 ±\pm 0.3610.2 ±\pm 1.1
LastDomain14.9 ±\pm 0.917.2 ±\pm 3.136.0 ±\pm 0.20.91 ±\pm 0.1825.8 ±\pm 0.610.3 ±\pm 0.169.1 ±\pm 0.7
IncFinetune16.7 ±\pm 3.410.1 ±\pm 0.834.0 ±\pm 0.30.83 ±\pm 0.0727.3 ±\pm 4.29.7 ±\pm 0.018.9 ±\pm 0.5
CDOT [ 32 ]9.3 ±\pm 1.014.2 ±\pm 1.034.1 ±\pm 0.00.94 ±\pm 0.1717.8 ±\pm 0.6––
CIDA [ 38 ]10.8 ±\pm 1.69.3 ±\pm 0.734.7 ±\pm 0.6–14.1 ±\pm 0.29.7 ±\pm 0.068.7 ±\pm 0.2
TABLE II : Performance comparison on DTDG datasets . Classification tasks report error rates (%), and regression tasks report MAE. “–” indicates that the method does not support the task.
Variant2-Moons-CTwitterCyclone
 Err. ↓\downarrowAUC ↑\uparrowMAE ↓\downarrow
Method-level variants
Fixed spectrum9.7 ±\pm 1.50.64 ±\pm 0.1117.2 ±\pm 0.7
Hard gating6.7 ±\pm 2.80.67 ±\pm 0.0918.7 ±\pm 0.3
Frozen gating2.8 ±\pm 0.30.70 ±\pm 0.0116.2 ±\pm 0.2
Objective-level variants
TABLE III : Ablation studies on representative CTDG benchmarks. We report error rate (%, ↓\downarrow ) for 2-Moons-C, AUC ( ↑\uparrow ) for Twitter, and MAE ( ↓\downarrow ) for Cyclone. The best performance in each column is in bold .
Fig. 4 : Qualitative spectral analysis on 2-Moons-C. (a) The parameter trajectory is decomposed into a smooth dominant component and a short-lived transient residual. (b) Learned modes are organized in the frequency–decay plane, where adaptive gates separate persistent modes from transient ones without fixed DFT bins or manual frequency thresholds. (c) Continuous-time inference extrapolates the dominant dynamics while freezing the transient component, producing a coherent future trajectory.
Fig. 5 : Qualitative comparison of decision boundaries learned on 2-Moons-C across six unseen future domains. Red curves denote learned decision boundaries alongside the corresponding classification error rates.
Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

DatasetPredictive model ( θ∈RD\theta\in\mathbb{R}^{D} )Latent encoder-decoder ( ϕ,ϕ−1\phi,\phi^{-1} )
2-Moons-CMLP: [2, 50, 50, 1] + ReLU + Sigmoidϕ\phi : [ D→1024→512→128→mD\to 1024\to 512\to 128\to m ] ϕ−1\phi^{-1} : [ m→128→512→1024→Dm\to 128\to 512\to 1024\to D ]
Rot-MNIST-C3×3\times [Conv(32/32/64, k=3k{=}3 ) + ReLU + MaxPool( 22 )] + MLP: [576, 128, 10] + LogSoftmax 
TwitterEmbedding Network + MLP [32, 128, 32, 1] + ReLU + Sigmoid 
Yearbook3×3\times [Conv(32/32/64, k=3k{=}3 ) + ReLU + MaxPool( 22 )] + MLP: [1024, 128, 32, 1] + ReLU + Sigmoid 
Cyclone4×4\times [Conv(32/32/64/64, k=3k{=}3 ) + ReLU + MaxPool( 22 )] + MLP: [1024, 128, 32, 1] + ReLU 
HouseMLP: [30, 400, 400, 1] + ReLU 
Appendix
TABLE IV : Detailed network architectures for the task predictive model and the uniform latent autoencoder ( ϕ,ϕ−1\phi,\phi^{-1} ) across all benchmark datasets. All autoencoder hidden layers employ ReLU activations.
Fig. 6 : t-SNE visualization on all 15 future domains of Rot-MNIST-C (colored by class). Each subfigure is one future domain with its step index and classification error.
Fig. 7 : t-SNE visualization on all 12 future domains of Yearbook (colored by class). Each subfigure is one future domain with its step index and classification error.
Fig. 8 : Hyperparameter sensitivity analysis of FreKoo++ with respect to objective loss weights α\alpha ( Lrec(C)\mathcal{L}_{\mathrm{rec}}^{(C)} ), β\beta ( Lfit(C)\mathcal{L}_{\mathrm{fit}}^{(C)} ), γ\gamma ( Rspec\mathcal{R}_{\mathrm{spec}} ), and δ\delta ( Rstab\mathcal{R}_{\mathrm{stab}} ) swept over {0.01,0.1,1.0,10,100}\{0.01,0.1,1.0,10,100\} across six benchmark datasets.
Fig. 9 : Sensitivity of the number of Koopman modes KK . The red star marks the adopted K=32K=32 .
More in this topic

Domain Generalization

Domain generalization develops models that perform reliably in new environments without training on target-domain data.

See all 369 papers

Explore similar work

CardsList
Oct 2, 2026

Adaptive Spectral-Koopman Dynamics Modeling for Temporal Domain Generalization

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 \href{}{https://anonymous.4open.science/r/Ada-Spec-K}.

Tengxue Zhang, Yu Ke, Yang Shu +4

East China Normal University · Chang’an University

Domain GeneralizationKoopman Operator Learning
AskPDF
Aug 21, 2026

Denoising the Future: Context-Aware Spectral Diffusion for Temporal Knowledge Graph Extrapolation

Temporal Knowledge Graph (TKG) extrapolation seeks to infer future facts from time-varying relational histories. Recent diffusion-based approaches improve uncertainty modeling through generative denoising, but their aggregated conditioning on subject histories may insufficiently distinguish query-specific evidence from non-salient historical facts, thereby diluting target-discriminative signals. To bridge this gap, we propose FreqDiff, a Frequency-aware Diffusion framework for TKG extrapolation. Specifically, FreqDiff formulates future object prediction as query-slot denoising and develops a dual-stream denoiser that integrates temporal dependency modeling with context-aware spectral calibration. The spectral branch synthesizes history-conditioned filters from learnable bases to adaptively re-calibrate denoising representations, while a frequency-domain regularizer is proposed to align the denoised target with the gold object in spectral space. Experiments on four public TKG benchmarks demonstrate that FreqDiff achieves state-of-the-art performance.

Yanglei Gan, Peng He, Run Lin +3

Southwest Minzu University, Chengdu, China · University of Electronic Science and Technology of China, Chengdu, China · Weixin Group, Tencent, Gunagzhou, China +1

Temporal KGsTemporal Link PredictionGraph DiffusionConditional Diffusion Models
AskPDF
Mar 27, 2026

Koopman Operator Identification of Model Parameter Trajectories for Temporal Domain Generalization (KOMET)

Parametric models deployed in non-stationary environments degrade as the underlying data distribution evolves over time (a phenomenon known as temporal domain drift). In the current work, we present KOMET (Koopman Operator identification of Model parameter Evolution under Temporal drift), a model-agnostic, data-driven framework that treats the sequence of trained parameter vectors as the trajectory of a nonlinear dynamical system and identifies its governing linear operator via Extended Dynamic Mode Decomposition (EDMD). A warm-start sequential training protocol enforces parameter-trajectory smoothness, and a Fourier-augmented observable dictionary exploits the periodic structure inherent in many real-world distribution drifts. Once identified, KOMET's Koopman operator predicts future parameter trajectories autonomously, without access to future labeled data, enabling zero-retraining adaptation at deployment. Evaluated on six datasets spanning rotating, oscillating, and expanding distribution geometries, KOMET achieves mean autonomous-rollout accuracies between 0.981 and 1.000 over 100 held-out time steps. Spectral and coupling analyses further reveal interpretable dynamical structure consistent with the geometry of the drifting decision boundary.

Randy C. Hoover, Jacob James, Paul May +1

Department of Electrical Engineering and Computer Science (EECS), South Dakota Mines, Rapid City, SD 57702, USA · Data Science and Engineering Program within EECS, South Dakota Mines, Rapid City, SD 57702, USA · Department of Mathematics, South Dakota Mines, Rapid City, SD 57702, USA

Extended Dynamic Mode DecompositionDrifting ModelsDomain GeneralizationLatent Dynamics ModelingKoopman Operator Learning
AskPDF