- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes TimeRift to separate background-conditioned dynamics with the deviations relative to the background. They think such separations will explain the similar temporal backgrounds will lead to different future trajectories. They will recombine the evolved deviation with the background for forecasting. They test their method on standard long-term forecasting benchmarks. The paper further provides ablation studies, temporal-background intervention experiments, efficiency and sensitivity studies to investigate the role of temporal conditioning.
Critical strengths
The paper is clearly written and the overall framework is easy to follow.
The authors provide a relatively broad empirical evaluation, including ablation studies, temporal-background interventions, sensitivity and efficiency analysis.
TimeRift is relatively lightweight and achieves competitive accuracy with modest memory and computational cost.
Critical weaknesses
1 The central idea to separate the background/residual with the temporal deviation has been well-explored in the previous methods like D-Linear, TQNet and CycleNet. The decomposition and recombination of the frequency are also mentioned in WaveletMixer and WPMixer. The remaining novel part might be spectral evolution. However, in Table 8, perturbing spectral evolution is only 0.69% to 3.11%, compared with around 180% for state calibration with recombination, which appears incremental.
2 The theorem and proposition are not informative. Proposition 1 is a direct result of the linearity of Fourier transform, while theorem 1 simply states the same local variation can transform differently with different background operators.
3 The empirical advantage is very small on each datasets, like 0.416 with 0.417 MSE and MAE on ETTh1, the 0.276 with 0.277 on ETTh2, also the almost equivalent performance on weather datasets. The authors report mean±std over five seeds for TimeRift, but the baseline results are presented without corresponding variances, and no pairwise significance tests or confidence intervals are provided. It is hard to believe the “superior overall forecasting performance over strong state-of-the-art baselines”.
4 One minor concern is under longer look-back lengths, TimeRift does not appear to maintain a clear advantage over the strongest baselines and the performance gap becomes quite small, even worse than TQNet in each setting in Figure 5b.
Questions for authors
1 How is TimeRift substantially different from existing decomposition and temporal-conditioning methods? Besides, why does w/o spectral evolution cause a larger performance drop than removing state calibration in the ablation study, while Table 8 shows only marginal sensitivity to perturbing spectral evolution compared with state calibration?
2 What additional insight do Proposition 1 and Theorem 1 provide beyond the model definition?
3 Are the small improvements over strong baselines statistically significant? How’s your method different from other frequency decomposition and recombined methods like WaveletMixer, WPMixer?
[1] Zhang, Zichi, et al. "WaveletMixer: a multi-resolution wavelets based MLP-mixer for multivariate long-term time series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 21. 2025.
[2] Murad, Md Mahmuddun Nabi, Mehmet Aktukmak, and Yasin Yilmaz. "Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025.
4 Frequency-domain methods often benefit from longer context windows, while the main experiments fix the look-back length to 96, which may limit the potential of TimeRift. The small gains of TimeRIFT on standard benchmark may be caused by the over-saturation. Some experiments on the broader benchmark like GIFT-EVAL can show the strength further.
Aksu, Taha, et al. "Gift-eval: A benchmark for general time series forecasting model evaluation." arXiv preprint arXiv:2410.10393 (2024).