acceptodds
Withdrawn from ICLR 2027

Value Lifting for Multivariate Time-Series Classification: From Fourier-Taylor Structure to Trainable High-Order Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing multivariate time-series classification (MTSC) models mainly improve temporal dependency modeling and cross-variable interaction, but they leave underexplored whether the scalar values these encoders produce benefit from an explicit, trainable transformation before classification. Here, we study this question through value lifting and instantiate it with Fourier-Taylor High-Order Lifting (FT-Lift), which converts intermediate values into a structured collection of odd- and even-order responses and learns how to combine them. Specifically, after coefficient relaxation, the Fourier-Taylor derivation provides a structured route to the high-order representation. The results show that FT-Lift improves the complete MTSC system, transfers to conventional MTSC backbones with architecture-dependent gains, achieves the best average rank among the compared lifting formulations, and provides task-dependent benefits from high-order value transformation. Overall, the evidence supports FT-Lift as a useful trainable value-lifting component for MTSC rather than a uniquely Fourier-specific function family.

Primary area:
learning on time series and dynamical systems
Keywords:
value lifting, multivariate time-series classification

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_C68KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates explicit value lifting for multivariate time-series classification (MTSC) and proposes Fourier-Taylor High-Order Lifting (FT-Lift), a trainable transformation of intermediate representations into structured high-order responses. The authors derive the proposed formulation from a truncated Fourier-Taylor expansion and subsequently relax its analytical coefficient constraints to obtain a learnable representation. The resulting module is integrated into VLNet and further evaluated on conventional CNN backbones. The paper conducts experiments on 26 UEA datasets, controlled comparisons of six lifting formulations on 10 datasets, cross-backbone transfer experiments, and synthetic high-order classification tasks. VLNet achieves an average accuracy of approximately 79.0% across the 26 datasets. The matched comparison shows that Relaxed-FT obtains the best average rank among the evaluated lifting formulations. However, its improvements over the no-lifting baseline and parameter-matched polynomial lifting are small and statistically inconclusive.

Critical strengths

S1. Clear theoretical characterization. The paper provides a careful derivation of FT-Lift and explicitly distinguishes the strict Fourier-Taylor formulation from its relaxed implementation. In particular, Proposition 2 clarifies that the relaxed formulation belongs to a polynomial function family rather than introducing additional Fourier-specific expressive capacity.

S2. Relatively comprehensive empirical analysis with meaningful controls. Beyond the broad 26-dataset benchmark, the authors investigate structural ablations, transfer to ResNet and FCN, parameter-matched lifting alternatives, and controlled high-order classification tasks. The inclusion of statistical tests and the acknowledgment of statistically inconclusive improvements are valuable.

Critical weaknesses

W1. The central methodological contribution is not sufficiently distinguished from existing polynomial representations. As established in Proposition 2, the fully relaxed FT-Lift formulation can be expressed as a polynomial mapping, with harmonic branches serving as a factorization of polynomial coefficients. Although this is theoretically well characterized, the paper does not establish a compelling practical advantage of this particular factorization. In Table 3, Relaxed-FT achieves an average accuracy of 0.7807, compared with 0.7786 for No Lift and 0.7773 for parameter-matched polynomial lifting. The corresponding differences are statistically inconclusive (p = 0.3750 and p = 0.2754, respectively).

W2. The experimental evidence supporting general effectiveness remains limited. The principal matched comparison covers only 10 UEA datasets, whereas the broader 26-dataset comparison relies on literature-reported baseline results obtained under heterogeneous experimental protocols. Furthermore, cross-backbone transfer exhibits substantially different improvements: 2.83 percentage points for ResNet but only 0.30 percentage points for FCN. These observations raise questions about the robustness and generality of the proposed component. The paper would benefit from stronger matched evaluations across more datasets and architectures, together with an analysis of computational overhead and performance variability.

Questions for authors

Q1. What practical advantage does the Fourier-Taylor-inspired factorization provide over a parameter-matched polynomial representation? Proposition 2 establishes that the relaxed formulation does not introduce a larger function family, while Table 3 reports statistically inconclusive differences. Can the authors provide additional evidence of improved optimization stability, convergence, computational efficiency, or generalization that specifically supports the proposed factorization?

Q2. How robust are the matched lifting comparisons across the full benchmark? Table 3 evaluates only 10 UEA datasets. Can the authors extend the same-protocol comparison among Relaxed-FT, No Lift, and parameter-matched polynomial lifting to all 26 datasets, or provide additional repeated-run evidence demonstrating that the observed improvements are reliable?

Q3. What explains the architecture-dependent effectiveness of FT-Lift? The reported improvement is considerably larger for ResNet than for FCN. Can the authors clarify which properties of the backbone or intermediate representations determine whether value lifting provides a meaningful benefit?

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

This paper studies explicit value lifting for multivariate time series classification. The proposed FT-Lift module applies a learnable combination of polynomial terms to intermediate features and is incorporated into a convolutional network, VLNet. Experiments evaluate classification performance, alternative lifting formulations, cross-backbone transfer, and sensitivity to polynomial order.

Critical strengths
  1. The paper presents FT-Lift, a trainable high-order representation module, and integrates it into a complete convolutional classification framework, VLNet.

  2. VLNet achieves the highest reported average accuracy and best average rank among the compared MTSC methods across 26 UEA datasets, demonstrating strong overall performance.

  3. The experiments provide broad empirical coverage through benchmark comparisons, structural ablations, controlled lifting comparisons, and sensitivity analyses. Improvements on ResNet and FCN further demonstrate the module’s applicability across different backbones.

Critical weaknesses
  1. The independent benefit of FT-Lift remains limited in the controlled comparison.

In Table 3, Relaxed-FT achieves 78.07% mean accuracy compared with 77.86% for No Lift, a difference of 0.21 percentage points that is not statistically significant. These results provide limited evidence that the proposed module consistently improves the underlying backbone.

  1. The value of explicit high-order transformation requires further clarification.

Several lifting variants in Table 3 perform below No Lift despite introducing additional trainable parameters. The parameter-matched polynomial control is useful, but comparisons with conventional convolutional or MLP modules under comparable parameter budgets would help distinguish the benefit of high-order structure from additional capacity.

  1. The scope of applicability remains unclear.

FT-Lift is a general feature transformation without an explicit classification-specific mechanism, while evaluation is restricted to classification. Although this matches the stated MTSC scope, experiments on forecasting or anomaly detection, or a clearer explanation of why the design particularly suits classification, would strengthen its motivation. Transfer to ResNet and FCN establishes cross-backbone applicability within classification, rather than cross-task generality.

  1. The relationship between Tables 1 and 3 could be clearer.

Both tables use the same ten datasets, but VLNet in Table 1 and Relaxed-FT in Table 3 report different average accuracies. Although Table 3 uses a repeated-run protocol, clarifying how the experimental settings and results correspond would improve transparency.

Questions for authors

Please refer to the Critical Weaknesses section above.

Reviewer_8qiNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper titled “ Value Lifting for Multivariate Time-Series Classification: From Fourier-Taylor Structure to Trainable High-Order Representations” proposes a value lifting technique for intermediate values of encoders to transform them into a structured collection of odd- and even-order responses and learns how to combine them. The proposed approach is an an element-wise transformation inserted between temporal feature extraction and final classification/pooling. The method is evaluated across 26 UEA benchmark datasets alongside synthetic Hermite polynomial tasks. Structural ablations, baseline comparisons, and cross-architecture transfers are used to demonstrate that this method provides task-dependent performance gains.

Critical strengths

The experimental section is detailed with No Lift, Exact Fourier, Strict-FT, Direct Polynomial, Parameter-Matched Polynomial using the exact same backbone, optimizer, and training protocol. They also conduct Holm corrections pointing out when differences fail to maintain statistical significance. Plugging FT-Lift into established architectures (vanilla ResNet and FCN) yields empirical gains across the 26 UEA datasets which shows some value in the proposed method.

Critical weaknesses

The biggest problem is that motivation for the paper is not convincing at all. In deep learning architectures, applying arbitrary element-wise transformations or additional non-linear functions to any intermediate activations can probably produce minor improvements in empirical resuts. There is no rigorous theoretical justification explaining why the intermediate value representations specifically require an explicit polynomial or Fourier-Taylor expansion before classification. The paper seems to try to show this empirically only, which is not a convincing argument. In some of the experiments shown, Relaxed-FT yields only a negligible numerical advantage over No Lift and Parameter-Matched Polynomial. So the empirical evidence is also marginal. Element wise operations are also limiting in that they cannot model multi-variate interaction which is often a hallmark of time series data.

Questions for authors

na

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

Related papers

Discussion (0)

Sign in to comment.