Machine learning algorithms for time-series

An important distinction in machine learning for time-series is the one between univariate and multivariate, in which algorithms are univariate, which means that they can only work with a single feature, or multi-variate, which means that they work with many features.

In univariate datasets, each case has a single series and a class label. Earlier models (classical modeling) focused on univariate datasets and applications. This is also reflected in the availability of datasets.

One of the most important repositories for time-series datasets, the UCR (University of California, Riverside) archive, which was released first in 2002, has provided a valuable resource for univariate time-series. It now contains about 120 datasets, but is lacking multivariate datasets. Furthermore, the M competitions (especially M3, 4, and 5) have a lot of available time-series datasets.

Multivariate time-series are datasets that have multiple feature...