Introduction to normaliseR

Purpose

normaliseR is a software package for R for rescaling numerical vectors or feature_calculations objects produced by the theft R package for computing time-series features.

Putting calculated feature vectors on an equal scale is crucial for any statistical or machine learning model as variables with high variance can adversely impact the model’s capacity to fit the data appropriately, learn appropriate weight values, or minimise a loss function. normaliseR includes function normalise (or normalize) to rescale either a whole feature_calculations object, or a single vector of values. The following normalisation methods are currently offered:

• z-score—"zScore"
• Sigmoid—"Sigmoid"
• Outlier-robust Sigmoid (credit to Ben Fulcher for creating the original MATLAB version) – "RobustSigmoid"
• Min-max—"MinMax"
• Maximum absolute—"MaxAbs"

normalise takes only three arguments:

1. data—either a feature_calculations object containing the raw feature matrix produced by theft::calculate_features or a numeric vector containing the values to be rescaled
2. norm_method—character denoting the rescaling/normalising method to apply. Can be one of "zScore", "Sigmoid", "RobustSigmoid", or "MinMax". Defaults to "zScore"
3. unit_int—Boolean whether to rescale into unit interval \([0,1]\) after applying normalisation method. Defaults to FALSE

Here is a simple example on a vector:

x <- rnorm(100)
normed <- normalise(x, norm_method = "zScore", unit_int = FALSE)

You can also access each individual rescaling function independently, though this affords you less overall control:

rs <- robustsigmoid_scaler(x)