arXiv cs.AI· Zelong Tian, Ming Jin, Bo Du, Shirui Pan·· 13 小时前AI 评分31
从频率视角重新审视时间序列 Tokenization：TimeFT 提出近无失真分词器
Rethinking Time Series Tokenization from a Frequency Perspective
AI 导读
针对现有时间序列 tokenization 策略引入频谱失真、使学习到的表示偏离趋势与周期性等真实模式的问题，研究者提出 TimeFT——一种基于频域划分、频率搬移与 Nyquist 采样的近无失真 tokenizer。该方法无参数、计算开销可忽略，可作为现有 tokenizer 的即插即用替代，在预测、分类与异常检测任务上均显著提升性能。该工作已被 NeurIPS 2026 接收。
正文 · 原文
Abstract:Tokenization that partitions time series into subsequences has become a foundational paradigm in modern time series modeling. However, we find that current tokenization strategies inherently introduce spectral distortions, forcing the learned representations to diverge from the true underlying patterns such as trend, periodicity, and seasonality, thereby impairing model performance. Through theoretical analysis from a frequency perspective, we derive the boundary conditions under which such distortions occur. To overcome this distortion, we propose TimeFT, a nearly distortion-free frequency-based tokenizer for time series data. TimeFT obtains tokens through frequency-domain partitioning, frequency shifting, and Nyquist sampling. The method is parameter-free, incurs negligible computational complexity, and can serve as a drop-in replacement for existing tokenizers. Our extensive experiments on diverse tasks such as forecasting, classification, and anomaly detection demonstrate that TimeFT consistently and significantly improves performance.
| Comments: | 40pages, Accepted by NeuIPS 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.28997 [cs.LG] |
| (or arXiv:2610.28997v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.28997 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zelong Tian [view email]
[v1]
Tue, 20 Oct 2026 23:51:24 UTC (724 KB)
来源：arXiv cs.AI · arxiv.org