Dask in machine learning originally consisted of a number of smaller libraries focused around particular sub-domains of machine learning.

While these special-purpose libraries were convenient for development, they were inconvenient for users who found the number of libraries daunting. The dask-ml project started as a combination of these that presented a single unified API and entry-point that mimicked Scikit-Learn. Afterwards additional algorithm development happened in the dask-ml library itself.

The pre-existing libraries are still valid and dask-ml defers to them for future development.