acceptodds
Withdrawn from ICLR 2027

CAW-KAN: Context-Aware Dual-Grid Wavelet KAN with Low-Rank Edge-Function Mixing for Long-Horizon Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series are rarely governed by a single simple relationship: once a variable is embedded into a latent representation, its future behavior depends on nonlinear interactions among the resulting latent coordinates. Kolmogorov-Arnold Networks (KANs) can model such interactions with learnable nonlinear edge functions, but giving every latent edge its own basis expansion requires a three-way coefficient tensor of size , which erodes efficiency as latent width or dictionary size grows. We present CAW-KAN, a channel-independent forecaster that separates temporal context extraction from nonlinear mixing in value space. For each observed variable, reversible instance normalization and a value embedding produce a -dimensional latent sequence, then a dense temporal convolution mixes across neighboring time steps and latent dimensions. A token-wise Wavelet-KAN evaluates a shared dictionary, built from two offset grids for complementary coverage of latent-value space, on the resulting scalar values; the grid centers and scales are currently fixed after initialization, with only the CP-factorized coefficients learned. To retain cross-feature edge functions without materializing the full coefficient tensor, CAW-KAN learns a low-rank CP factorization of these coefficients. Across 12 datasets and four forecast horizons, CAW-KAN attains the lowest error or ties for it in 59 of 96 MSE/MAE cells. Resource measurements show that CAW-KAN operates with significantly fewer parameters than other state-of-the-art models. Here, "context-aware" refers only to the local temporal representation fed into the Wavelet-KAN, not to a context gate or a temporal wavelet transform. The source code and implementation details are publicly available at https://github.com/tftpbe/CAW-KAN.

Primary area:
learning on time series and dynamical systems
Keywords:
Long-Horizon Time Series Forecasting, Kolmogorov–Arnold Networks (KAN), Wavelet-KAN, Dual-Grid Wavelet Dictionary, Low-Rank Tensor Factorization, CP Decomposition, Temporal Context Modeling

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_1qE7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

CAW-KAN is a dual grid wavelet + KAN based model for time series forecasting. This is a channel independent forecaster which first applies revin and conv 1D layers for embedding, then dense temporal Conv 1D + Wavelet-KAN blocks sequentially, finally a projector to obtain final predictions. KAN uses a fixed dual-grid dictionary in latent value space with CP-factored coefficients. Authors reported best/tied results in multiple datasets and baselines.

Critical strengths
  1. Paper includes experiments on multiple forecast lengths, seeds and datasets which increases the breadth of evaluation and statistical significance.
  2. A very simple way of explaining the architecture and experiments are provided with matched loss results.
Critical weaknesses
  1. Very limited architectural novelty: The normalisation (RevIN), Conv1D based mixing, channel independent processing all the methods are from well known previous papers like TimeKAN etc. Authors just added a Wavelet-KAN blocks which is very incremental
  2. Limited performance gains than baselines: Authors have shown handpicked and single seed results to show in main paper, Table 10 clearly says performance gains are very less compared to baselines like HaKAN and TimeKAN. In addition even in table 1, proposed method is better only in few settings.
Questions for authors
  1. Why Without Wavelet-KAN block ablation is not available in paper ?
  2. Authors mentioned, CAW-KAN trains with Log-Cosh and the baselines with MSE is a fair comparison, how is that a fair comparison ?
  3. Why a handpicked and single seed results are only shown in main paper ?
  4. in ablation KAN to MLP (Table 2), is that a parameter matched MLP ?
Reviewer_hYxbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CAW-KAN is a channel-independent long-horizon forecaster combining RevIN, convolutional context mixing, a per-token Wavelet-KAN and a linear head. Its main choices are a dual-grid wavelet dictionary with fixed centers and scales and a directly trained CP low-rank coefficient tensor. The paper gives a closed-form overlap calibration and a parameter/MAC analysis, and reports best or tied-best results in 59 of 96 cells over 12 datasets. The appendix adds a same-loss, multi-seed comparison on the four ETT datasets.

Critical strengths

1 A simple design with a clear parameterization. The fixed dual grid and CP factors make a concrete, lightweight model, and the overlap calibration and parameter/MAC counts are easy to follow. 2 The same-loss, multi-seed experiment is real positive evidence. CAW-KAN is strictly best in 14 of 32 cells in Table 10, mostly on ETTm1 and ETTm2. This does not show overall superiority, but I think it shows practical value on some datasets.

Critical weaknesses
  1. The main experiments do not isolate the benefit of the design. CAW-KAN uses Log-Cosh while the baselines use MSE, and baseline tuning budgets are unclear. Table 10 helps but covers only four ETT datasets and three baselines, with gains that vary by dataset. The paper does not report d_model, and the public scripts use a small width and dictionary, so the claimed CP benefit at large width is never tested against a same-backbone baseline. The promised latency and memory measurements are also missing.
  2. Anomalies in the results reduce my trust in the main comparison. In Tables 5 and 7, DLinear and iTransformer match in 15 of 16 three-decimal values across the Aotizhongxin and Changping stations. Some best values are also mismarked: Aotizhongxin-336 MSE marks 0.704 although HaKAN has 0.703, and Exchange-96 MAE marks 0.213 although TimeMixer has 0.210. The repetition alone does not prove an error, but without run logs I can't be confident in these comparisons.
  3. The model stacks known pieces: RevIN with channel independence, convolutional mixing, a fixed-basis wavelet KAN, CP low-rank factorization and a linear head. Low-rank KANs, basis-function swaps and channel-independent forecasters all have precedents, and closely related KAN forecasting work such as RMoK, TKAN and FastKAN is not discussed. The linear head alone holds about 91% of the reported parameters on ETTh1, so the KAN part is a small share of the model. The name also oversells the design, since the abstract itself says "context-aware" only refers to the local temporal representation fed into the KAN.
Questions for authors

Why are 15 of 16 baseline values identical across the two air-quality stations? Per-station data paths, run logs and unrounded metrics would help, and I would also like to know why the nominal default configuration gives different values in Table 8. Where do the baseline runs come from and how were they tuned, and does the pattern in Table 10 hold on the other datasets with a common loss and multiple seeds? On the same backbone, how much do the fixed dual grid and CP each contribute compared with a full-rank version, and what are the promised latency and memory numbers? Lastly, which commit matches the submitted experiments, given that the public code has a tanh saturation not in the equations and lacks the RevIN clamp? Why do the wavelet scales in Fig. 2 differ from the Eq. 13 calibration?

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

This paper proposes CAW-KAN, a lightweight time-series forecasting framework that combines dense temporal convolution, a dual-grid Wavelet-KAN, and CP-factorized edge-function coefficients. The dual-grid dictionary aims to provide complementary coverage of the latent-value space, while CP factorization reduces the parameter and computational costs of the KAN layer. The model is evaluated on 12 time-series datasets across four forecasting horizons, with comparisons against conventional forecasting architectures and existing KAN-based methods. The results are promising, particularly in terms of parameter efficiency and competitive forecasting accuracy.

Critical strengths
  1. A reasonable architecture for efficient forecasting. The combination of temporal context extraction and nonlinear latent-feature mixing is well motivated. CP factorization provides a principled way to reduce the parameterization cost of the edge-function coefficient tensor while retaining cross-feature interactions within each observed variable.

  2. Broad empirical evaluation. The paper evaluates 12 datasets at four forecasting horizons and includes recent KAN-based forecasting baselines, such as TimeKAN and HaKAN, as well as conventional linear, convolutional, recurrent, Transformer-based, and lightweight foundation-model approaches.

  3. Useful mathematical and empirical analyses. The paper provides parameter and operation-count analyses, derives a closed-form initialization rule for the Mexican-Hat wavelets, and investigates component contributions and hyperparameter sensitivity. The derivations appear generally consistent with their stated assumptions, although they do not establish that the proposed design is optimal for forecasting.

Critical weaknesses
  1. The proposed framework combines several established directions in KAN research: wavelet-based edge functions, low-rank parameterization, and temporal-context modeling. For example, Wav-KAN has already explored wavelet-based KAN functions, while TimeKAN and HaKAN investigate alternative KAN architectures for time-series forecasting. Low-rank and parameter-reduced KAN approaches, such as LoTRA and PRKAN, also demonstrate that reducing the parameterization cost of KANs is an established research direction. I do not consider the combination of these components inherently problematic; a useful integration can constitute a contribution. However, the paper should more clearly distinguish its specific innovations from these existing ideas, particularly the dual-grid initialization and the direct CP factorization of the Wavelet-KAN coefficient tensor.

  2. The paper compares CAW-KAN with TimeKAN and HaKAN, which is a strength. Nevertheless, the related literature on KAN-based forecasting and parameter-efficient KAN architectures is not fully represented in the experiments. In particular, KFS and HiKAN are relevant recent approaches that could strengthen the comparison. These methods are not identical to CAW-KAN, and their inclusion should be subject to compatible evaluation protocols. Still, comparing against them, or at least discussing their methodological differences more systematically, would help clarify the advantages of the proposed framework. Similarly, adapting existing parameter-reduced or low-rank KAN strategies to the same forecasting architecture would provide a more direct assessment of the CP parameterization. I regard this as a relatively minor weakness compared with the need for controlled component-level validation.

  3. The paper includes ablations of single-grid versus dual-grid configurations and examines different CP ranks. These experiments are useful, but the evidence does not yet fully establish the independent benefits of the proposed design choices. For the dual-grid dictionary, a more informative comparison would examine alternative basis configurations under matched parameter budgets, such as a denser single grid or learnable centers and scales. This would clarify whether the observed gains arise specifically from the proposed dual-grid arrangement rather than simply from the allocation and resolution of basis functions. For CP factorization, the paper demonstrates parameter reduction and reports sensitivity to the factorization rank. However, the parameter–accuracy trade-off is not sufficiently characterized. I would like to see forecasting error and parameter count reported together across a broader range of ranks, ideally with comparisons against a dense coefficient tensor and an alternative low-rank parameterization. Such experiments would establish whether CP factorization achieves a favorable efficiency–accuracy trade-off and identify the rank regimes in which it is most beneficial.

  4. In the primary experiments, CAW-KAN is trained with Log-Cosh loss, whereas the baselines use MSE, and the main results use a single initialization seed. The appendix includes additional multi-seed comparisons using a common Log-Cosh objective, but these cover only a subset of models and the four ETT datasets. This makes it difficult to attribute the overall performance differences exclusively to the proposed architecture. In particular, the ablations suggest that changing the loss function can have a noticeable effect on forecasting accuracy. More systematic comparisons using matched loss functions and multiple random seeds would strengthen the evidence for the independent contributions of the dual-grid dictionary and CP-factorized Wavelet-KAN.

Questions for authors
  1. How does the dual-grid initialization compare with alternative wavelet-grid designs, such as denser single grids or learnable centers and scales, when parameter budgets are controlled?

  2. Can the authors provide a parameter–accuracy trade-off across different CP ranks and compare CP factorization with a dense coefficient tensor or another low-rank KAN parameterization?

  3. Can the authors clarify the differences between CAW-KAN and recent KAN-based forecasting approaches, particularly KFS and HiKAN, and provide additional comparisons where the evaluation protocols are compatible?

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

Related papers

Discussion (0)

Sign in to comment.