Algorithms
changepoynt implements several change point detection approaches from subspace estimation, density-ratio estimation, and time-series segmentation.
| Algorithm | Source | Status |
|---|---|---|
| SST | Ide, 2005 | Stable |
| Fast SST | Weber et al., 2025 | Stable |
| IKA-SST | Ide, 2007 | Stable |
| Fast IKA-SST | Weber et al., 2025 | Stable |
| Multivariate (IKA-)SST (MSST) | Stable | |
| ESST | Boelter & Weber et al., 2025 | Stable |
| Multivariate ESST (MESST) | Stable | |
| RuLSIF | Liu et al. | Stable |
| uLSIF | Liu et al. | Stable |
| KLIEP | Liu et al. | Planned |
| ClaSP | Ermshaus et al. | Deactivated |
| FLUSS | Gharghabi et al. | Stable |
| FLOSS | Gharghabi et al. | Unavailable |
| BOCPD | Adams et al. | Experimental |
| MovingWindow | Wu & Keogh | Stable baseline |
| ZERO | van den Burg & Williams | Stable baseline |
| Subspace Identification | Kawahara et al. | Unavailable |
| TESST | Experimental |
Choosing a Starting Point
For most one-dimensional signals, start with ESST or SST. If runtime becomes an issue for large window sizes, enable the accelerated Hankel implementation where supported.
For parameter choices, runtime trade-offs, and decomposition methods for SST, ESST, MSST, and MESST, see Tuning Subspace Methods.
For streaming-style matrix-profile segmentation, consider FLOSS. For offline matrix-profile segmentation, consider FLUSS.
Density-ratio methods such as ULSIF and RuLSIF are useful when the change is best described as a distribution shift between neighboring windows.
User Guides
The user guides group practical advice by method family. This keeps the algorithm table useful as a catalog while allowing each guide to discuss related methods together.
| Method family | Guide |
|---|---|
| Subspace methods | Tuning SST, ESST, MSST, and MESST |
| Fast subspace methods | Fast SST and IKA-SST |