TimeTok: Hierarchical Tokenization for Granularity-Controllable Time-Series Generation
TimeTok serves as a cohesive system for generating time-series data with adjustable granularity (GC-TSG). It employs a hierarchical tokenization method to convert time series into sequential tokens that range from broad to detailed temporal granularity. The autoregressive generation method functions across different levels of granularity, creating token blocks that are then transformed into continuous time series. By regulating the quantity of token blocks, users can precisely manage the level of detail in the output. Experimental results indicate that TimeTok performs exceptionally well in GC-TSG tasks, allowing for generation at any desired granularity, whether starting from coarser input or from the beginning.
Key facts
- TimeTok introduces granularity-controllable time-series generation (GC-TSG).
- Hierarchical tokenization maps time series into coarse-to-fine token sequences.
- Autoregressive generation produces token blocks decoded into continuous time series.
- Controlling token blocks provides explicit control over output detail.
- TimeTok generates time series at any target granularity from coarser input or from scratch.
- Experiments demonstrate TimeTok's effectiveness at GC-TSG.
Entities
—