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

2026-09-24

AI Summary

This study addresses the loss of temporal evolution order and cross-sample feature semantic inconsistency when applying tabular foundation models to time series classification. Treating representation design as an independent task, this work proposes a novel paradigm that preserves local temporal transitions while maintaining shared feature definitions. By introducing shared switching dynamics and mechanism codebook learning, local dynamic operators become directly comparable across samples, enabling efficient inference via a frozen in-context tabular foundation model. The proposed method achieves the highest average accuracy on evaluation benchmarks, outperforming the strongest baseline by 4.47%, while demonstrating robust performance in few-shot scenarios.

Abstract

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.

Problem

• time series classification
• tabular foundation models
• sequence-to-tabular representation
• temporal transitions
• shared feature definition

Innovation

• Tabular Foundation Models
• Time Series Classification
• Shared Switching Dynamics
• Regime Codebook
• In-Context Learning