Update KMeans init docstring to include 'random' option (#1020)

The docstring for the dask_ml.cluster.KMeans class currently omits 'random'
as a valid option for the init parameter. However, this option is fully
implemented in the underlying k_init function and serves as a critical
scalable alternative to the default 'k-means||', which can overwhelm the
scheduler on large datasets.

Author: sergeyklay

dask_ml/cluster/k_means.py

@@ -36,19 +36,24 @@ class KMeans(TransformerMixin, BaseEstimator):
     ----------
     n_clusters : int, default 8
         Number of clusters to end up with
-    init : {'k-means||', 'k-means++' or ndarray}
+    init : {'k-means||', 'k-means++', 'random' or ndarray}
         Method for center initialization, defaults to 'k-means||'.
 
-        'k-means||' : selects the the gg
+        'k-means||' : Selects initial cluster centers using a scalable
+        variant of k-means++. See the notes for more details.
 
-        'k-means++' : selects the initial cluster centers in a smart way
+        'k-means++' : Selects the initial cluster centers in a smart way
         to speed up convergence. Uses scikit-learn's implementation.
 
         .. warning::
 
            If using ``'k-means++'``, the entire dataset will be read into
            memory at once.
 
+        'random' : Selects `n_clusters` random rows from the input data for
+        the initial centroids. Use `n_init` to run multiple random
+        initializations for more robust results.
+
         An array of shape (n_clusters, n_features) can be used to give
         an explicit starting point