KBinsDiscretizer

KBinsDiscretizer is a feature transformer in scikit-learn that discretizes continuous features into a fixed number of bins. It converts numeric attributes into categorical or ordinal representations, which can be useful for models that benefit from discretized inputs or for preprocessing within pipelines.

The transformer supports per-feature bin specification through the n_bins parameter, which can be an int or an array-like specifying the number of bins for each feature. The encode option determines how the discretized data is returned: 'ordinal' provides integer bin indices, while 'onehot' yields a sparse one-hot encoded matrix and 'onehot-dense' returns a dense array. The strategy option defines how bin edges are computed: 'uniform' creates equal-width bins, 'quantile' aims for approximately equal numbers of samples per bin, and 'kmeans' uses k-means clustering to define bin edges.

During fitting, KBinsDiscretizer computes the bin edges for each feature according to the selected strategy. Transform then assigns each value to a bin and applies the chosen encoding. The resulting feature representation varies in shape: ordinal encoding returns a matrix with the same number of columns as features, while one-hot encodings expand to a number of columns equal to the sum of the bin counts across features.

Typical usage involves importing KBinsDiscretizer, configuring n_bins, encode, and strategy, and applying fit_transform to the data. For example, disc = KBinsDiscretizer(n_bins=[5, 3], encode='ordinal', strategy='uniform'); X_disc = disc.fit_transform(X).

Notes include considerations about the impact of the chosen strategy on bin widths and robustness, as well as potential increases in dimensionality with one-hot encodings. The transformer integrates with scikit-learn’s pipelines and cross-validation.