API Reference / preprocessing /

discretization/

Binning continuous features into discrete intervals.

Turning a numeric column into a small number of bins. This is what algorithms that expect categorical input need, and it can help elsewhere by letting a model express a non-linear relationship it could not otherwise fit.

Transformers

  • EqualWidthDiscretizer: Bins of equal range. Simple, but a skewed
feature leaves most bins nearly empty.
  • QuantileDiscretizer: Bins holding equal numbers of samples. Robust to
skew, at the cost of uneven bin widths.
  • MDLDiscretizer: Chooses cut points by minimum description length,
using the target to place boundaries where the class actually changes, and deciding the number of bins for you.

Notes

MDL is supervised: it reads y. Fit it on the training split only — fitting on everything chooses boundaries informed by the test labels and inflates the score.

Discretising always discards information. It is worth it when the algorithm requires it or the relationship is genuinely non-monotonic, not by default.

Modules

equal_frequency

QuantileDiscretizer transformer.

1 class · 0 functions
equal_width

EqualWidthDiscretizer transformer.

1 class · 0 functions
mdl

MDLDiscretizer transformer.

1 class · 0 functions