Shared switching dynamics improve time series classification accuracy
SwitchPFN: Shared Switching Dynamics for Frozen In-Context Time Series Classification
Machine Learning
Summary
Classifying time series data can be tricky because capturing the order and meaning of changes over time is hard. The authors present SwitchPFN, a method that learns shared patterns and codes from training data to keep temporal information consistent across different sequences. This helps make local changes more comparable and improves classification accuracy. Their approach outperforms previous methods, especially when there is limited labeled data available.
What this means in practice
- •For data science teams: Improve time series classification models by using shared dynamic representations across samples to boost accuracy on sequential data.
- •For machine learning engineers: Develop models that maintain consistent temporal features across sequences for better performance in limited labeled data scenarios.
Authors
Zhenyi Zhu, Jacqueline Pang, Peilin Shen, Tianyi Song, Tingwei Zhang, Keyi Hu, Kangjun Yin, Shiwei Pu, Yingbo Zhou, Chen Shao
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.