We recently merged a discretizing transformer into the discrete branch (see diff between that branch and master).

The current discretizer there makes fixed width bins based on each feature's input range. While we could do quantile bins by preceding this in a pipeline with a QuantileTransformer, this is inefficient. Quantile bins is quite a common approach to discretizing.

Another approach to discretizing performs KMeans clustering (or perhaps even another clusterer or mixture model) in each feature.

We should either have separate classes for these discretization approaches, or provide a strategy= parameter to the current proposed discretizer.

To dear contributor: Make sure to submit a pull request to the discrete branch.