acceptodds
Withdrawn from ICLR 2027

Learning Locality-Aware Basis Functions via Spline-Wavelet Kolmogorov-Arnold Networks for Functional Data Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Functional data analysis (FDA) is widely used to model continuous processes from discrete observations across scientific domains. However, conventional FDA methods often rely on fixed basis families, leading to inefficient representations when discriminative patterns are concentrated within localized regions. To address this limitation, this work proposes Spline-Wavelet KAN for Adaptive Basis Learning (SKAB), a neural method that learns locality-aware basis functions for functional data classification. Motivated by an observed structural correspondence between FDA and Kolmogorov-Arnold Networks (KANs), SKAB formulates a KAN as a Coordinate-to-Basis Generator CBG that maps each functional-domain coordinate to basis functions. To better capture localized patterns, Spline-Wavelet KAN is introduced to parameterize the CBG using compactly supported B-spline wavelet atoms as edge functions. Each functional observation is projected onto the learned bases to obtain a coefficient representation for downstream prediction. Experiments on synthetic and real-world datasets demonstrate that SKAB achieves competitive classification performance, with particular advantages for data containing localized discriminative patterns. The code is available at this link.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Kolmogorov-Arnold Networks, Functional Data Analysis, Representation Learning, Classification, Spline Wavelets

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_aPQPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes SKAB, a supervised adaptive basis-learning method for functional data classification. A coordinate-input Spline-Wavelet KAN generates shared basis functions, with edge functions combining compactly supported wavelet atoms with learnable centers and scales and a SiLU branch. The proposed CSDS procedure initializes the first-layer atom dictionary using coordinate-wise Fisher scores computed from the training data. Each curve is converted into coefficients through a differentiable ridge least-squares projection, followed by an MLP classifier. The basis generator and classifier are trained jointly using the classification objective. Evaluation includes a synthetic task constructed from localized atoms, six real datasets, and ablations of the parameterization, initialization, and dictionary size.

Critical strengths
  1. The motivation is clear. When class differences are concentrated in narrow regions, adapting the location and scale of the representation using supervision is a reasonable alternative to retaining only dominant global variation. SKAB implements this idea through trainable local atoms and shared functional representations.

  2. The paper includes targeted component comparisons. Beyond final classification accuracy, it replaces the basis-generator parameterization, removes supervised initialization, and visualizes localized responses. These analyses help characterize the model's behavior rather than reporting predictive scores alone.

Critical weaknesses
  1. The central performance advantage is not established through sufficiently fair and relevant comparisons. The paper emphasizes the large improvement on EthanolLevel, but Appendix B reports that simple supervised db4 coefficient selection with the same MLP already achieves 0.8176±0.01230.8176 \pm 0.0123 at K=8K=8, compared with 0.8160±0.01480.8160 \pm 0.0148 for SKAB. The protocols differ, as the authors acknowledge, so these numbers do not establish statistical equivalence. Nevertheless, they substantially weaken the attribution of the main-table improvement to the proposed KAN basis-learning mechanism. Fixing representation-based methods at K=6K=6 also does not establish their relative performance after appropriate capacity selection. In the preliminary Phoneme sweep, B-splines at K=12K=12 already match SKAB's mean accuracy in that table. More importantly, Appendix A defers principled center-scale dictionary-size selection to future work while reporting substantial configuration sensitivity. For a method whose current performance depends on this choice, a reproducible validation-based selection procedure and corresponding fair baseline comparisons are necessary evidence for the present claims, rather than an optional future extension.

  2. The mathematical properties of individual atoms are insufficiently connected to the locality and interpretability of the final representation. Compact support and vanishing moments belong to the mother wavelet and its translated and scaled atoms. Subsequent layers operate on hidden representations, and the network retains global SiLU branches. These properties therefore do not automatically establish locality of the final basis functions in the original functional coordinate or suppression of local linear backgrounds by the resulting coefficients. The coefficient computation additionally involves the inverse basis Gram matrix, which can mix information across components. Energy and component-contribution plots describe particular fitted representations, but do not establish stable recovery of discriminative regions. Conditioning of the basis matrix, redundancy among learned functions, and localization stability across random seeds are insufficiently assessed. The evidence therefore supports an empirical inductive bias from localized atoms, but not stronger conclusions about structural recovery or interpretability.

Questions for authors
  1. Under identical data splits, random seeds, prediction heads, and validation-based model selection, does SKAB significantly outperform supervised db4 coefficient selection? Please include validation-selected KK and computational cost, rather than restricting the comparison to six-dimensional representations.

  2. How were the number of centers QQ, number of scales MM, and representation dimension KK selected in the main experiments? In particular, what justified the different center count for EthanolLevel? Under a common search procedure using training and validation data only, how much of SKAB's advantage over appropriately tuned fixed-basis methods remains?

  3. Which properties of the final basis functions or coefficients follow from the mother wavelet's compact support and vanishing moments after multilayer nonlinear composition, global SiLU branches, and ridge projection? Please distinguish atom-level properties, empirical localization, and mathematical guarantees for the final representation.

  4. Can the authors report the spectrum or condition number of the learned Gram matrix and the consistency of identified regions across random seeds? On the synthetic task, can localization recovery be evaluated quantitatively against the known ground truth to establish whether the interpretations suggested by the energy maps are stable?

Reviewer_fKxDOpenReview ↗
Rating
2
Confidence
4
Informativeness
1(−1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SKAB, a locality-aware basis learning framework for functional data classification. It uses a KAN-based coordinate-to-basis generator with compactly supported spline-wavelet atoms and a Fisher-score-based initialization strategy (CSDS). Input curves are projected onto the learned bases via ridge-regularised least squares, followed by an MLP classifier. Experiments on one synthetic and six real-world datasets show competitive performance, with SKAB achieving the highest mean accuracy on five of the six real-world datasets. Ablation studies further evaluate the proposed components.

Critical strengths
  1. The proposed design is well motivated for localised functional patterns. The use of spline wavelets with learnable centers and scales makes sense for classification tasks where discriminative features are localized. The Fisher-score-based initialisation is also a straightforward way to place the wavelet atoms in informative regions.

  2. The paper includes useful comparisons and ablation studies. The paper compares SKAB against a range of classical and neural FDA methods, along with several KAN and time-series models. The ablations on different CBG parameterisations and CSDS help show which parts of the method contribute to its performance.

Critical weaknesses
  1. The methodological novelty and added value over existing approaches remain unclear. Task-adaptive basis learning and wavelet-based KANs have been explored in prior work, while the KAN–FDA connection in Eqs. (2)–(4) is not new in itself. The main contribution appears to be the combination of spline-wavelet KAN edges with Fisher-score-based initialisation. However, the results in Appendix B suggest that simpler supervised wavelet methods can achieve comparable performance. I am therefore not fully convinced that the additional complexity of the learned CBG is justified.

  2. The experimental reporting raises concerns about reliability. The main text and appendices contain several inconsistencies, including different scale parameterisations, B-spline KAN results, while SoftShape is omitted from the main comparison. Appendix A also reports preliminary sensitivity analyses with validation-based hyperparameter selection still pending. These issues make the results difficult to assess.

Questions for authors
  1. What are the main methodological advantages of SKAB over the existing adaptive basis learning methods discussed in the Introduction and Related Work?

  2. Appendix B shows that supervised db4 wavelet selection achieves accuracy comparable to SKAB on EthanolLevel. Could the authors compare these methods under the same experimental settings to isolate the benefit of learning the basis functions?

  3. How stable is the subspace spanned by the learned basis functions across different training runs? Do the learned representations consistently capture similar discriminative regions?

  4. How were the number of basis functions and the center-scale dictionary size (Q, M) selected across datasets? In particular, was the choice of Q = 4 for EthanolLevel based solely on validation data?

  5. The evaluation includes six real-world datasets, which cover only part of the range of functional classification problems. Could the authors include more functional classification datasets, particularly those with global or mixed discriminative patterns, to better assess how well SKAB generalises?

  6. How well do the final learned basis functions preserve locality after passing through the multi-layer CBG? Appendix H.1 notes that the final bases “need not be compactly supported with respect to the original coordinate”. Could the authors provide a measure of this?

  7. Section 3.3 and Appendix describe different scale parameterisations and bounds. Which configuration was used in the reported experiments?

  8. Could you please clarify the AI Use Statement? It states that AI was not used to implement the methods, but later mentions AI-assisted code generation and editing. It also contains an incomplete sentence and refers to surveys or interviews, which are not part of this study. The authors should revise the statement to clarify how AI was used.

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

The authors propose SKAB for basis expansions for functional data that combine KANs with wavelets. They motivate their work before providing their approach alongside empirical work.

Critical strengths

The usage of KANs in ML represents an interesting alternative to the now standard MLPs used extensively in deep learning. The idea of exploring how to leverage them with functional data seems like an interesting topic.

Critical weaknesses

The biggest issue is that for 1D functions (inputs are 1D), the KAN basically buys you nothing. The whole motivation for the Kolmogorov–Arnold theorem is that you can express complex multivariate functions using only 1D functions, which is a genuinely interesting idea. But the paper is about 1D functional data, meaning you don't get really anything from referencing KANs. Equations 3–4 are basically saying nothing: 'you can express a 1D function as a sum of 1D functions'.

The other issue is that KAN doesn't actually give you a recipe for how to build your functions. The authors went with wavelets, which is completely reasonable. But given my first point, their method boils down to using wavelets to express the functional data. And there are many many papers that already do that. They cite Chang et al (2014), but there are now many papers that use wavelets with functional data, going back to at least the early 2000s (e.g., Brown, Fearn & Vannucci, JASA 2001; Morris & Carroll, JRSS-B 2006; Wang, Ray & Mallick, JASA 2007; Zhu, Brown & Morris, Biometrics 2012), and even earlier in signal processing (Saito & Coifman's local discriminant bases, 1995). This includes supervised approaches that select discriminative wavelet coefficients for classification (e.g., Vannucci, Sha & Brown 2005), so the fact that the atoms are learned with the predictive objective is not new on its own. The authors would need to really explain why their method contributes to that body of literature (which is largely absent from the present paper).

I want to stress though that I think the authors have an interesting idea/angle with exploring KANs with functional data, I just don't think this use case makes much sense.

Questions for authors

See weaknesses.

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

This paper proposes to a method named SKAB, which is for functional data classification. It aims to improve the performance on datasets where class-discriminative information is concentrated in a narrow region of the domain. The method uses a single input KAN with the coordinate t as input and outputs a set of K basis functions. Each edge function of the KAN is parameterised by compactly supported order-2 Chui-Wang spline wavelet atoms. The centers and scales of the first-layer atoms are initialised from the training set Fisher score, so that more atoms are placed in discriminative regions.Each curve is projected onto the learned basis by ridge regression, and the resulting coefficients are fed into an MLP to predict the label. The whole model is trained with Adam and stopping based on the validation loss. Experiments by a synthetic dataset and six real-world datasets show that SKAB achieves competitive accuracy. The authors claim that the method has particular advantages for data containing localised discriminative patterns.

Critical strengths

The main strength of this paper is that it provides an interpretable, task-adaptive framework for learning localised bases, and validates it on a synthetic dataset with ground-truth atoms. The method uses a KAN as a basis generator, which generates bases directly from coordinates, and parameterises its edge functions with compactly supported order-2 Chui-Wang spline-wavelet atoms. As a result, the learned bases can be directly visualised. In the synthetic experiment, where the discriminative regions are known, Fig. 3 shows that both the Fisher-initialised centres and the trained bases align with the true atoms. This is consequential because it provides direct evidence of interpretability, beyond higher accuracy, which supports the central claim of the paper. However, this evidence is limited to the single-layer architecture used for the synthetic data.

Critical weaknesses

The paper claims that the SKAB method achieve higher accuracy than others on on five of the six datasets. The question is the form only give a accuracy test result under a single fixed train-test split.The experiments only change the random seed uniformly and repeat the runs ten times, without changing the data split, and no significance test is given.In the Tecator dataset, the number of test samples is very small, only 43, so one sample corresponds to about 2.3 percentage points, the 1% accuracy difference is relatively small. The Medflies dataset also has only 107 test samples, which also has a large influence on the accuracy. According to Appendix G, regarding the parameters of each method, the paper states in Table 10 that “for comparison methods that use different settings, we follow the configurations provided in their original implementations.” The tuning of parameters may also have a large effect on the accuracy. The main text in Eq. 3, using a describe as “a scalar-input two-layer KAN with hidden width K”, which suggests that the bases are given directly by the first-layer edge functions. However, in table 11, the KAN network mentioned as [1,64,64,K], which implies it might be a multi-layer KAN network is using in experiment.As the synthetic experiment uses a single [1,K] layer “On DJ-MIXTURE, SKAB uses a single [1,K] SWK layer without hidden CBG layers. ” This difference is not discussed in the main text.

Questions for authors

1.Could the authors report results over multiple random train-validation-test splits (or cross-validation), especially for Tecator and Medflies, whose test sets contain only 43 and 107 samples?

2.Does SKAB use a single [1,K] layer on the synthetic data but a [1,64,64,K] architecture on the real-world datasets? Could the authors report the results of the single-layer version on the real-world datasets?

Reviewer_5UM6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SKAB for functional data classification. A coordinate-input KAN generates shared basis functions using spline-wavelet atoms, whose initial locations and scales are determined from training-set Fisher scores. Each observation is represented through ridge projection onto the learned functions, followed by an MLP classifier. Experiments on one synthetic and six real datasets examine classification performance and the contribution of individual components.

Critical strengths

Learning the locations and scales of wavelet atoms is a reasonable approach when class differences are concentrated in small regions of the functional domain. The supervised initialization provides a direct way to incorporate this information into basis learning.

Critical weaknesses

1.The locality of individual atoms does not directly imply locality of the final representation.

  1. The benefit of the KAN basis generator over simpler supervised wavelet methods is not sufficiently established.
Questions for authors
  1. Under the same splits, prediction head, and validation-based choice of representation dimension, how does SKAB compare with supervised db4 selection and a single-layer adaptive wavelet generator?

  2. Can the authors quantify localization of the final basis functions and its stability across runs, particularly for the deeper architecture used real data?

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

Related papers

Discussion (0)

Sign in to comment.