Preprocessing

dask_ml.preprocessing contains some scikit-learn style transformers that can be used in Pipelines to perform various data transformations as part of the model fitting process. These transformers will work well on dask collections (dask.array, dask.dataframe), NumPy arrays, or pandas dataframes. They’ll fit and transform in parallel.

Scikit-Learn Clones

Some of the transformers are (mostly) drop-in replacements for their scikit-learn counterparts.

MinMaxScaler([feature_range, copy]) Transforms features by scaling each feature to a given range.
QuantileTransformer([n_quantiles, …]) Transforms features using quantile information.
RobustScaler([with_centering, with_scaling, …]) Scale features using statistics that are robust to outliers.
StandardScaler([copy, with_mean, with_std]) Standardize features by removing the mean and scaling to unit variance
LabelEncoder Encode labels with value between 0 and n_classes-1.

These can be used just like the scikit-learn versions, except that:

  1. They operate on dask collections in parallel
  2. .transform will return a dask.array or dask.dataframe when the input is a dask collection

See sklearn.preprocessing for more information about any particular transformer.

Additional Tranformers

Other transformers are specific to dask-ml.

Categorizer([categories, columns]) Transform columns of a DataFrame to categorical dtype.
DummyEncoder([columns, drop_first]) Dummy (one-hot) encode categorical columns.
OrdinalEncoder([columns]) Ordinal (integer) encode categorical columns.

Both dask_ml.preprocessing.Categorizer and dask_ml.preprocessing.DummyEncoder deal with converting non-numeric data to numeric data. They are useful as a preprocessing step in a pipeline where you start with heterogenous data (a mix of numeric and non-numeric), but the estimator requires all numeric data.

In this toy example, we use a dataset with two columns. 'A' is numeric and 'B' contains text data. We make a small pipeline to

  1. Categorize the text data
  2. Dummy encode the categorical data
  3. Fit a linear regression
In [1]: from dask_ml.preprocessing import Categorizer, DummyEncoder

In [2]: from sklearn.linear_model import LogisticRegression

In [3]: from sklearn.pipeline import make_pipeline

In [4]: import pandas as pd

In [5]: import dask.dataframe as dd

In [6]: df = pd.DataFrame({"A": [1, 2, 1, 2], "B": ["a", "b", "c", "c"]})

In [7]: X = dd.from_pandas(df, npartitions=2)

In [8]: y = dd.from_pandas(pd.Series([0, 1, 1, 0]), npartitions=2)

In [9]: pipe = make_pipeline(
   ...:    Categorizer(),
   ...:    DummyEncoder(),
   ...:    LogisticRegression()
   ...: )
   ...: 

In [10]: pipe.fit(X, y)
Out[10]: 
Pipeline(memory=None,
     steps=[('categorizer', Categorizer(categories=None, columns=None)), ('dummyencoder', DummyEncoder(columns=None, drop_first=False)), ('logisticregression', LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,
          intercept_scaling=1, max_iter=100, multi_class='ovr', n_jobs=1,
          penalty='l2', random_state=None, solver='liblinear', tol=0.0001,
          verbose=0, warm_start=False))])

Categorizer will convert a subset of the columns in X to categorical dtype (see here for more about how pandas handles categorical data). By default, it converts all the object dtype columns.

DummyEncoder will dummy (or one-hot) encode the dataset. This replaces a categorical column with multiple columns, where the values are either 0 or 1, depending on whether the value in the original.

In [11]: df['B']
Out[11]: 
0    a
1    b
2    c
3    c
Name: B, dtype: object

In [12]: pd.get_dummies(df['B'])
Out[12]: 
   a  b  c
0  1  0  0
1  0  1  0
2  0  0  1
3  0  0  1

Wherever the original was 'a', the transformed now has a 1 in the a column and a 0 everywhere else.

Why was the Categorizizer step necessary? Why couldn’t we operate directly on the object (string) dtype column? Doing this would be fragile, especially when using dask.dataframe, since the shape of the output would depend on the values present. For example, suppose that we just saw the first two rows in the training, and the last two rows in the tests datasets. Then, when training, our transformed columns would be:

In [13]: pd.get_dummies(df.loc[[0, 1], 'B'])
Out[13]: 
   a  b
0  1  0
1  0  1

while on the test dataset, they would be:

In [14]: pd.get_dummies(df.loc[[2, 3], 'B'])
Out[14]: 
   c
2  1
3  1

Which is incorrect! The columns don’t match.

When we categorize the data, we can be confident that all the possible values have been specified, so the output shape no longer depends on the values in the whatever subset of the data we currently see. Instead, it depends on the categories, which are identical in all the subsets.