Creating and selecting features for a time series
In the previous recipe, we automatically extracted several hundred features from a time series variable using tsfresh. If we have more than one time series variable, we can easily end up with a dataset that contains thousands of features.
When we create classification and regression models to solve real-life problems, we often want our models to take a small number of relevant features as input to produce their predictions. Simpler models have many advantages. First, their output is easier to interpret for the end users of the models. Second, simpler models are cheaper to store and faster to train. They also return their outputs faster.
The tsfresh library provides a highly parallel feature selection algorithm based on non-parametric statistical hypothesis tests, which can be executed at the back of the feature creation procedure to quickly remove irrelevant features. The feature selection procedure utilizes different tests...