Time Series Segmentation

A particularly interesting problem which we will study in this notebook is time series segmentation (TSS).

TSS aims at discovering regions of a time series that are semantically dissimilar to neighboring regions. TSS is an important technology, as it allows to infer properties of the underlying system by analysis of the measurements, because shifts from one segment to another are often caused by state changes in the process being monitored, such as a transition from one operational state to another or the onset of an unusual event. Change point detection (CPD) is the task of finding such shifts in the underlying signal, and a segmentation is an ordered sequence of change points.

The following image shows two examples of (a) heartbeats recorded from a patient with different medical conditions, and (b) a person walking, jogging and running. The time series contain three visually very distinct segments, which we also denoted by colors and annotated by the underlying state.

The task of any TSS algorithm is to find these segments.

We currently have two notebooks on segmentation

• ClaSP (Classification Score Profile) Segmentation

• Heteregeneous Intrinsic Dimensionality Algorithm (Hidalgo) Segmentation

---

Generated using nbsphinx. The Jupyter notebook can be found here.