TableVectorizer#

class skrub.TableVectorizer(*, cardinality_threshold=40, low_cardinality=OneHotEncoder(drop='if_binary', dtype='float32', handle_unknown='ignore', sparse_output=False), high_cardinality=GapEncoder(n_components=30), numeric=PassThrough(), datetime=DatetimeEncoder(), specific_transformers=(), n_jobs=None)[source]#

Transform a dataframe to a numeric (vectorized) representation.

Applies a different transformation to each of several kinds of columns:

  • numeric: floats, integers, and booleans.

  • datetime: datetimes and dates.

  • low_cardinality: string and categorical columns with a count of unique values smaller than a given threshold (40 by default). Category encoding schemes such as one-hot encoding, ordinal encoding etc. are typically appropriate for columns with few unique values.

  • high_cardinality: string and categorical columns with many unique values, such as free-form text. Such columns have so many distinct values that it is not possible to assign a distinct representation to each: the dimension would be too large and there would be too few examples of each category. Representations designed for text, such as topic modelling (GapEncoder) or locality-sensitive hashing (MinHash) are more appropriate.

Note

Transformations are applied independently on each column. A different transformer instance is used for each column separately; multivariate transformations are therefore not supported.

The transformer for each kind of column can be configured with the corresponding parameter. A transformer is expected to be a compatible scikit-learn transformer. Special-cased strings "drop" and "passthrough" are accepted as well, to indicate to drop the columns or to pass them through untransformed, respectively.

Additionally, it is possible to specify transformers for specific columns, overriding the categorization described above. This is done by providing a list of pairs (transformer, list_of_columns) as the specific_transformers parameter.

Note

The specific_transformers parameter will be removed in a future version of skrub, when better utilities for building complex pipelines are introduced.

Parameters:
cardinality_thresholdint, default=40

String and categorical columns with a number of unique values strictly smaller than this threshold are handled by the transformer low_cardinality, the rest are handled by the transformer high_cardinality.

low_cardinalitytransformer, “passthrough” or “drop”, default=OneHotEncoder instance

The transformer for string or categorical columns with strictly fewer than cardinality_threshold unique values. By default, we use a OneHotEncoder that ignores unknown categories and drop one of the transformed columns if the feature contains only 2 categories.

high_cardinalitytransformer, “passthrough” or “drop”, default=GapEncoder instance

The transformer for string or categorical columns with at least cardinality_threshold unique values. The default is a GapEncoder with 30 components (30 output columns for each input).

numerictransformer, “passthrough” or “drop”, default=”passthrough”

The transformer for numeric columns (floats, ints, booleans).

datetimetransformer, “passthrough” or “drop”, default=DatetimeEncoder instance

The transformer for date and datetime columns. By default, we use a DatetimeEncoder.

specific_transformerslist of (transformer, list of column names) pairs, default=()

Override the categories above for the given columns and force using the specified transformer. This disables any preprocessing usually done by the TableVectorizer; the columns are passed to the transformer without any modification. A column is not allowed to appear twice in specific_transformers. Using specific_transformers provides similar functionality to what is offered by scikit-learn’s ColumnTransformer.

n_jobsint, default=None

Number of jobs to run in parallel. None means 1 unless in a joblib parallel_backend context. -1 means using all processors.

Attributes:
transformers_dict

Maps the name of each column to the fitted transformer that was applied to it.

column_to_kind_dict

Maps each column name to the kind ("high_cardinality", "low_cardinality", "specific", etc.) it was assigned.

kind_to_columns_dict

The reverse of column_to_kind_: maps each kind of column ("high_cardinality", "low_cardinality", etc.) to a list of column names. For example kind_to_columns['datetime'] contains the names of all datetime columns.

input_to_outputs_dict

Maps the name of each input column to the names of the corresponding output columns.

output_to_input_dict

The reverse of input_to_outputs_: maps the name of each output column to the name of the column in the input dataframe from which it was derived.

all_processing_steps_dict

Maps the name of each column to a list of all the processing steps that were applied to it. Those steps may include some pre-processing transformations such as converting strings to datetimes or numbers, the main transformer (e.g. the DatetimeEncoder), and a post-processing step casting the main transformer’s output to numpy.float32. See the “Examples” section below for details.

feature_names_in_list of str

The names of the input columns, after applying some cleaning (casting all column names to strings and deduplication).

n_features_in_int

The number of input columns.

all_outputs_list of str

The names of the output columns.

See also

tabular_learner

A function that accepts a scikit-learn estimator and creates a pipeline combining a TableVectorizer, optional missing value imputation and the provided estimator.

Examples

>>> from skrub import TableVectorizer
>>> import pandas as pd
>>> df = pd.DataFrame({
...     'A': ['one', 'two', 'two', 'three'],
...     'B': ['02/02/2024', '23/02/2024', '12/03/2024', '13/03/2024'],
...     'C': ['1.5', 'N/A', '12.2', 'N/A'],
... })
>>> df
       A           B     C
0    one  02/02/2024   1.5
1    two  23/02/2024   N/A
2    two  12/03/2024  12.2
3  three  13/03/2024   N/A
>>> df.dtypes
A    object
B    object
C    object
dtype: object
>>> vectorizer = TableVectorizer()
>>> vectorizer.fit_transform(df)
   A_one  A_three  A_two  B_year  B_month  B_day  B_total_seconds     C
0    1.0      0.0    0.0  2024.0      2.0    2.0     1.706832e+09   1.5
1    0.0      0.0    1.0  2024.0      2.0   23.0     1.708646e+09   NaN
2    0.0      0.0    1.0  2024.0      3.0   12.0     1.710202e+09  12.2
3    0.0      1.0    0.0  2024.0      3.0   13.0     1.710288e+09   NaN

We can inspect which outputs were created from a given column in the input dataframe:

>>> vectorizer.input_to_outputs_['B']
['B_year', 'B_month', 'B_day', 'B_total_seconds']

and the reverse mapping:

>>> vectorizer.output_to_input_['B_total_seconds']
'B'

We can also see the encoder that was applied to a given column:

>>> vectorizer.transformers_['B']
DatetimeEncoder()
>>> vectorizer.transformers_['A']
OneHotEncoder(drop='if_binary', dtype='float32', handle_unknown='ignore',
              sparse_output=False)
>>> vectorizer.transformers_['A'].categories_
[array(['one', 'three', 'two'], dtype=object)]

We can see the columns grouped by the kind of encoder that was applied to them:

>>> vectorizer.kind_to_columns_
{'numeric': ['C'], 'datetime': ['B'], 'low_cardinality': ['A'], 'high_cardinality': [], 'specific': []}

As well as the reverse mapping (from each column to its kind):

>>> vectorizer.column_to_kind_
{'C': 'numeric', 'B': 'datetime', 'A': 'low_cardinality'}

Before applying the main transformer, the TableVectorizer applies several preprocessing steps, for example to detect numbers or dates that are represented as strings. Moreover, a final post-processing step is applied to all non-categorical columns in the encoder’s output to cast them to float32. We can inspect all the processing steps that were applied to a given column:

>>> vectorizer.all_processing_steps_['B']
[CleanNullStrings(), ToDatetime(), DatetimeEncoder(), {'B_day': ToFloat32(), 'B_month': ToFloat32(), ...}]

Note that as the encoder (DatetimeEncoder() above) produces multiple columns, the last processing step is not described by a single transformer like the previous ones but by a mapping from column name to transformer.

all_processing_steps_ is useful to inspect the details of the choices made by the TableVectorizer during preprocessing, for example:

>>> vectorizer.all_processing_steps_['B'][1]
ToDatetime()
>>> _.format_
'%d/%m/%Y'

Transformers are applied separately to each column

The TableVectorizer vectorizes each column separately – a different transformer is applied to each column; multivariate transformers are not allowed.

>>> df_1 = pd.DataFrame(dict(A=['one', 'two'], B=['three', 'four']))
>>> vectorizer = TableVectorizer().fit(df_1)
>>> vectorizer.transformers_['A'] is not vectorizer.transformers_['B']
True
>>> vectorizer.transformers_['A'].categories_
[array(['one', 'two'], dtype=object)]
>>> vectorizer.transformers_['B'].categories_
[array(['four', 'three'], dtype=object)]

Overriding the transformer for specific columns

We can also provide transformers for specific columns. In that case the provided transformer has full control over the associated columns; no other processing is applied to those columns. A column cannot appear twice in the specific_transformers.

Note

This functionality is likely to be removed in a future version of the TableVectorizer.

The overrides are provided as a list of pairs: (transformer, list_of_column_names).

>>> from sklearn.preprocessing import OrdinalEncoder
>>> vectorizer = TableVectorizer(
...     specific_transformers=[('drop', ['A']), (OrdinalEncoder(), ['B'])]
... )
>>> df
       A           B     C
0    one  02/02/2024   1.5
1    two  23/02/2024   N/A
2    two  12/03/2024  12.2
3  three  13/03/2024   N/A
>>> vectorizer.fit_transform(df)
     B     C
0  0.0   1.5
1  3.0   NaN
2  1.0  12.2
3  2.0   NaN

Here the column ‘A’ has been dropped and the column ‘B’ has been passed to the OrdinalEncoder (instead of the default choice which would have been DatetimeEncoder).

We can see that ‘A’ and ‘B’ are now treated as ‘specific’ columns:

>>> vectorizer.column_to_kind_
{'C': 'numeric', 'A': 'specific', 'B': 'specific'}

Preprocessing and postprocessing steps are not applied to columns appearing in specific_columns. For example ‘B’ has not gone through ToDatetime():

>>> vectorizer.all_processing_steps_
{'A': [Drop()], 'B': [OrdinalEncoder()], 'C': [CleanNullStrings(), ToFloat32(), PassThrough(), {'C': ToFloat32()}]}

Specifying several specific_transformers for the same column is not allowed.

>>> vectorizer = TableVectorizer(
...     specific_transformers=[('passthrough', ['A', 'B']), ('drop', ['A'])]
... )
>>> vectorizer.fit_transform(df)
Traceback (most recent call last):
    ...
ValueError: Column 'A' used twice in 'specific_transformers', at indices 0 and 1.

Methods

fit(X[, y])

Fit transformer.

fit_transform(X[, y])

Fit transformer and transform dataframe.

get_feature_names_out()

Return the column names of the output of transform as a list of strings.

get_metadata_routing()

Get metadata routing of this object.

get_params([deep])

Get parameters for this estimator.

set_output(*[, transform])

Set output container.

set_params(**params)

Set the parameters of this estimator.

transform(X)

Transform dataframe.

fit(X, y=None)[source]#

Fit transformer.

Parameters:
Xdataframe of shape (n_samples, n_features)

Input data to transform.

yarray_like, shape (n_samples,) or (n_samples, n_outputs) or None, default=None

Target values for supervised learning (None for unsupervised transformations).

Returns:
selfTableVectorizer

The fitted estimator.

fit_transform(X, y=None)[source]#

Fit transformer and transform dataframe.

Parameters:
Xdataframe of shape (n_samples, n_features)

Input data to transform.

yarray_like of shape (n_samples,) or (n_samples, n_outputs) or None, default=None

Target values for supervised learning (None for unsupervised transformations).

Returns:
dataframe

The transformed input.

get_feature_names_out()[source]#

Return the column names of the output of transform as a list of strings.

Returns:
list of strings

The column names.

get_metadata_routing()[source]#

Get metadata routing of this object.

Please check User Guide on how the routing mechanism works.

Returns:
routingMetadataRequest

A MetadataRequest encapsulating routing information.

get_params(deep=True)[source]#

Get parameters for this estimator.

Parameters:
deepbool, default=True

If True, will return the parameters for this estimator and contained subobjects that are estimators.

Returns:
paramsdict

Parameter names mapped to their values.

set_output(*, transform=None)[source]#

Set output container.

See Introducing the set_output API for an example on how to use the API.

Parameters:
transform{“default”, “pandas”, “polars”}, default=None

Configure output of transform and fit_transform.

  • “default”: Default output format of a transformer

  • “pandas”: DataFrame output

  • “polars”: Polars output

  • None: Transform configuration is unchanged

Added in version 1.4: “polars” option was added.

Returns:
selfestimator instance

Estimator instance.

set_params(**params)[source]#

Set the parameters of this estimator.

The method works on simple estimators as well as on nested objects (such as Pipeline). The latter have parameters of the form <component>__<parameter> so that it’s possible to update each component of a nested object.

Parameters:
**paramsdict

Estimator parameters.

Returns:
selfestimator instance

Estimator instance.

transform(X)[source]#

Transform dataframe.

Parameters:
Xdataframe of shape (n_samples, n_features)

Input data to transform.

Returns:
dataframe

The transformed input.