Tailoring feature creation to different time series
The tsfresh library extracts many features based on the time series characteristics and distribution, such as their correlation properties, stationarity, entropy, and non-linear time series analysis functions, which decompose the time series signal through, for example, Fourier or wavelet transformations. Depending on the nature of the time series, some of these transformations make more sense than others. For example, wavelength decomposition methods can make sense for time series resulting from signals or sensors but are unsuitable for time series representing sales or stock prices.
In this recipe, we will discuss how to optimize the tsfresh feature extraction procedure to create specific features for each time series, and then use these features to predict office occupancy.
How to do it...
The tsfresh library accesses the methods that will be used to create features through a dictionary that contains the method’...
