PaperlayerView exact revision on arXiv

Source-linked AI summary

SwitchPFN: Shared Switching Dynamics for Frozen In-Context Time Series Classification

Zhenyi Zhu, Jacqueline Pang, Peilin Shen, Tianyi Song, Tingwei Zhang, Keyi Hu, Kangjun Yin, Shiwei Pu, Yingbo Zhou, Chen Shao

arXiv:2609.29814v1cs.LG

TL;DR

Time-series representations for frozen tabular foundation models must retain local temporal order while keeping feature meanings comparable across sequences. SwitchPFN uses a shared projection and regime codebook to encode local dynamics and transitions. Across eight UEA datasets, it achieves 82.63 ± 0.45% mean accuracy, 3.54 percentage points above matched MultiRocket + TabPFN.

  • Problem

    For frozen tabular foundation models, time-series representations need to encode local-pattern order and changes in evolution while keeping columns comparable across sequences.

    Source: §1 · p. 1

  • Method

    SwitchPFN uses a shared projection and regime codebook to represent local dynamics and regime transitions for a frozen TabPFN predictor.

    Source: §1 · p. 2, §2.1 · p. 2

  • Results

    82.63 ± 0.45% mean accuracy on eight UEA datasets was 3.54 percentage points above matched MultiRocket + TabPFN at 79.09 ± 0.40%.

    Source: §1 · p. 2, §3.2 · p. 4

  • Takeaways & Limitations

    The authors conclude that shared coordinates, a common regime codebook, and class-specific residual features provide complementary information.

    Source: §4 · p. 4

    The formulation assumes labeled sequences with finite time points and channels, with labels drawn from C classes.

    Source: §2 · p. 2

Abstract

from arXiv · show

Tabular foundation models (TFMs) provide a promising route to time-series classification, but their effectiveness depends on how sequential data are converted into tabular representations. Existing representations face two challenges: global aggregation can lose the order of temporal evolution, while features computed in independently fitted coordinate systems may not have consistent meanings across sequences. We therefore view representation design for TFMs as a problem in its own right: the representation should preserve local temporal transitions while maintaining a shared feature definition across samples. We propose SwitchPFN, which learns a shared projection and regime codebook from the training sequences, making local dynamic operators and transition features directly comparable across samples. Across the evaluated benchmarks, SwitchPFN achieves the highest mean accuracy among the evaluated methods, improving over the strongest baseline by 4.47% relatively. Ablation studies, parameter sensitivity analyses, and reduced-training-data experiments further examine the contributions of the representation, its main design choices, and its behavior when labeled data are limited.

1. INTRODUCTION

The introduction frames time-series representation for frozen tabular foundation models as needing both temporal order and feature meanings shared across sequences. SwitchPFN addresses these requirements with shared coordinates and a regime codebook.

1. INTRODUCTION

  • Time-series classification assigns labels from signal values and temporal changes, with applications including ECG arrhythmia detection, sleep staging, and astronomical transient classification.Prior approaches include convolutional and periodic-pattern models, as well as transformer and contrastive pretraining.Source: §1 · p. 1
  • The representation problem is to retain the order of local patterns while giving each feature column the same meaning across sequences.Frequency alone does not determine order, and operator coefficients can vary with the coordinate system used to express a sequence.Source: §1 · p. 1
  • On eight UEA datasets, SwitchPFN achieved 82.63% mean test accuracy, 3.54 percentage points higher than the second-best model; ablation, sensitivity, and label-efficiency experiments assessed its components, parameter choices, and data efficiency.Source: §1 · p. 2
  • SwitchPFN represents local-pattern order and regime changes with shared projection and codebook features in a fixed-budget table for TabPFN.The design combines switching dynamics, class residuals, ordered paths, and signal summaries.Source: §1 · p. 2

2. METHODOLOGY

SwitchPFN maps sequences into shared features for frozen TabPFN prediction, combining local dynamics, class-specific residuals, ordered paths, and signal summaries.

2.1. Shared Coordinate System

  • A shared PCA basis projects normalized signals and first differences into common coordinates, so operator coefficients refer to the same directions across examples.The center and projection are determined from training data.Source: §2.1 · p. 2

2.2. Local Dynamics and Shared Regimes

  • Local windows define affine evolution features, which are encoded against shared regime centers and summarized through operator moments, occupancy, dwell times, and lagged transitions.Window encodings combine operator deviations, intercepts, residuals, and velocity summaries; ridge regularization adapts to input scale.Source: §2.2 · p. 2

2.3. Class-Specific Dynamics Banks

  • Class-specific affine banks yield residual features; training rows exclude their own class contribution, while query residuals use complete training banks.Five features per class summarize horizon scores, temporal residual variance, Gaussian negative log-likelihood, and the margin to the best competing class.Source: §2.3 · pp. 2–3

2.4. Feature Construction and Frozen Prediction

  • Four feature groups combine switching dynamics, class residuals, path signatures, and signal summaries; hybrid selection standardizes a fixed-width feature vector for frozen TabPFN prediction.A reserved set of columns is followed by columns selected using training-label Fisher scores; training rows form the context, and query rows use full class banks.Source: §2.4 · p. 3

3. EXPERIMENTS

Experiments on eight UEA datasets compare SwitchPFN with classical, neural, convolution-based, and frozen tabular predictors. SwitchPFN has the highest mean test accuracy, while ablations, sensitivity analyses, and reduced-label experiments examine its components and training-data behavior.

3.1. Experimental Setup

  • Evaluation uses official UEA train/test splits across eight tasks spanning five signal domains, 3–61 channels, 2–26 classes, and varied sequence lengths.Baselines include classical, neural, convolution-plus-Ridge, and frozen tabular predictor families.Source: §3.1 · p. 3

3.2. Main Results

  • 82.63 ± 0.45% mean test accuracy gives SwitchPFN a 3.54 pp gain over matched MultiRocket + TabPFN (79.09 ± 0.40%), or a 4.47% relative improvement.SwitchPFN achieves the highest mean among the evaluated methods.Source: Figure 3 · p. 4, §3.2 · p. 4
  • SwitchPFN beats the matched control on seven of eight datasets, with at least one win in each of five signal domains.Excluding EC, it wins six of seven datasets, with a mean gain of 1.81 pp.Source: §3.2 · p. 4

3.3. Model Analysis

  • 79.01% is the frozen reference OOF accuracy; increasing r_0 to 24 gives 79.62%, while retaining only the first 1,024 columns gives 69.34%.The analysis uses training-only out-of-fold validation; accuracy is more sensitive to feature selection than rank or budget.Source: Table 2 · p. 4, §3.3 · p. 4
  • Removing transitions loses 0.68 pp; other ablations report losses of 0.84–9.06 pp, including 9.06 pp for structured-only features.A single regime loses 1.47 pp, removing class residuals loses 0.93 pp, a random shared projection loses 0.84 pp, and replacing TabPFN with Ridge loses 5.67 pp; generic-only features lose 1.76 pp.Source: Figure 4 · p. 4, §3.3 · p. 4
  • With 25%, 50%, and 100% of the remaining training pool, SwitchPFN reaches validation accuracies of 68.45%, 75.40%, and 79.38%, respectively.Its gains over MiniRocket + TabPFN are 3.32, 6.66, and 6.38 pp at those fractions; the advantage persists with a quarter of the labeled pool.Source: Figure 5 · p. 4, §3.3 · p. 4

4. CONCLUSION AND FUTURE WORK

The conclusion presents SwitchPFN as a shared-space representation for local dynamics and regime transitions, evaluated with a frozen TabPFN predictor.

  • Experiments on eight UEA datasets show improved mean accuracy and effective performance with limited labels; future work targets broader domains, independent feature-selection validation, and lower feature-construction cost.The representation combines shared coordinates, a common regime codebook, and class-specific residual features.Source: §4 · p. 4

DECLARATION OF AI USE

The authors report that ChatGPT was used only for language editing, and that they reviewed and approved all suggested revisions.

COMPLIANCE WITH ETHICAL STANDARDS

The work uses publicly available benchmark datasets and reports no collection of new human-participant data.

Loading 2609.29814v1…