analiza aiqbench
This research addresses a fundamental flaw in current time series modeling where standard tokenization methods distort spectral features like trends and seasonality. The authors propose TimeFT, a frequency-domain tokenizer that eliminates these distortions to significantly improve performance across forecasting, classification, and anomaly detection tasks. By operating as a parameter-free drop-in replacement, this method offers a more faithful representation of underlying data patterns for advanced AI systems.
Rethinking Time Series Tokenization from a Frequency Perspective
arXiv:2610.28997v1 Announce Type: new 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.
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.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2610.28997 [cs.LG]
(or arXiv:2610.28997v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.28997 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 229 cuvinte
Subiecte conexe
Beyond Uniform Tokens: Adaptive Compression for Time Series Language Models
arXiv · 12 iun. 2026
TRE: Training-Free Hallucination Detection for Diffusion Language Models
arXiv · 28 iul. 2026
Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models
arXiv · 30 sept. 2026
DiFA: Dual Evidence Fusion and Aggregation for Token-Level Text Anomaly Detection
arXiv · 22 sept. 2026