Unsupervised methods for time-series
The main difference between time-series and other types of data is the dependence on the temporal axis; a correlation structure at one point t1 could have very different information to the same structure at point t2. Time-series often contain lots of noise and have high dimensionality.
To reduce the noise and decrease the dimensionality, dimensionality reduction, wavelet analysis, or signal processing techniques such as filtering, for example, Fourier decomposition, can be applied. This is often at the basis of anomaly detection or CPD, the techniques we are discussing in this chapter. We'll discuss drift detection in Chapter 8, Online Methods for Time-Series.
We'll be talking in detail about anomalies and change points, and it might be helpful to see an illustration of how anomalies and change points can look like. In the article "Social tipping dynamics for stabilizing Earth's climate by 2050" by Ilona Otto and...