Add cross_validate helper

[TomAugspurger]

sklearn.model_selection.cross_validate fits and scores several models over some CV splits of data.

http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.cross_validate.html

Users can currently do this distributed on a cluster with

import dask_ml.joblib

with joblib.parallel_backend('dask', scatter=[X, y]):
    sklearn.model_selection.cross_validate(estimator, X, y)

Why not do that for them by defining a dask_ml.model_selection.cross_validate that does the parallel_backend and scattering for them?

[TomAugspurger]

We would do the same for cross_val_score and cross_val_predict.

[TomAugspurger]

Why not do that for them...?

One reason might be if we want to implement a cross_validate that's specifically designed to work with with distributed data (RAM bound). This little wrapper would only be helpful for CPU bound workloads. But we could presumably dispatch to our implementation when given dask collections.

[gglanzani]

But we could presumably dispatch to our implementation when given dask collections.

@TomAugspurger Is there a dask implementation that works specifically for dask collections? Or is it in the making/will happen some day?

[TomAugspurger]

Is there a dask implementation that works specifically for dask collections?

Dask-ML currently only implements ShuffleSplit for dask arrays (not dataframes).

Opened https://github.com/dask/dask-ml/issues/269 for implementing additional splitters.

---

For this issue, I would say that dask_ml.model_selection.cross_validate could start off with checking if is_dask_collection(X) or is_dask_collection(y): raise NotImplementedError with a nice message.