Merge pull request #177 from jnothman/pipeline-minimal

explain_weights in Pipelines: minimal version

docs/source/autodocs/eli5.rst

@@ -13,3 +13,5 @@ The following functions are exposed to a top level, e.g.
 .. autofunction:: eli5.show_weights
 
 .. autofunction:: eli5.show_prediction
+
+.. autofunction:: eli5.transform_feature_names

docs/source/libraries/sklearn.rst

@@ -195,6 +195,23 @@ is independent.
 .. _ExtraTreesClassifier: http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesClassifier.html#sklearn.ensemble.ExtraTreesClassifier
 .. _ExtraTreesRegressor: http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesRegressor.html#sklearn.ensemble.ExtraTreesRegressor
 
+.. _sklearn-pipelines:
+
+Transformation pipelines
+------------------------
+
+:func:`eli5.explain_weights` can be applied to a scikit-learn Pipeline_ as
+long as:
+
+* ``explain_weights`` is supported for the final step of the Pipeline
+* :func:`eli5.transform_feature_names` is supported for all preceding steps
+  of the Pipeline. singledispatch_ can be used to register
+  ``transform_feature_names`` for transformer classes not handled (yet) by ELI5
+  or to override the default implementation.
+
+.. _Pipeline: http://scikit-learn.org/stable/modules/generated/sklearn.pipeline.Pipeline.html#sklearn.pipeline.Pipeline
+.. _singledispatch: https://pypi.python.org/pypi/singledispatch
+
 Reversing hashing trick
 -----------------------
 

eli5/__init__.py

@@ -6,6 +6,7 @@
 from .formatters import format_as_html, format_html_styles, format_as_text
 from .explain import explain_weights, explain_prediction
 from .sklearn import explain_weights_sklearn, explain_prediction_sklearn
+from .transform import transform_feature_names
 
 
 try:

eli5/sklearn/__init__.py

@@ -17,3 +17,4 @@
     FeatureUnhasher,
     invert_hashing_and_fit,
 )
+from . import transform as _

eli5/sklearn/explain_weights.py

@@ -5,6 +5,7 @@
 import numpy as np  # type: ignore
 
 from sklearn.base import BaseEstimator, RegressorMixin  # type: ignore
+from sklearn.pipeline import Pipeline  # type: ignore
 from sklearn.linear_model import (   # type: ignore
     ElasticNet,  # includes Lasso, MultiTaskElasticNet, etc.
     ElasticNetCV,
@@ -61,6 +62,7 @@
     get_default_target_names,
 )
 from eli5.explain import explain_weights
+from eli5.transform import transform_feature_names
 from eli5._feature_importances import (
     get_feature_importances_filtered,
     get_feature_importance_explanation,
@@ -422,3 +424,17 @@ def _features(target_id):
             method='linear model',
             is_regression=True,
         )
+
+
+@register(Pipeline)
+def explain_weights_pipeline(estimator, feature_names=None, **kwargs):
+    last_estimator = estimator.steps[-1][1]
+    transform_pipeline = Pipeline(estimator.steps[:-1])
+    if 'vec' in kwargs:
+        feature_names = get_feature_names(feature_names, vec=kwargs.pop('vec'))
+    feature_names = transform_feature_names(transform_pipeline, feature_names)
+    out = explain_weights(last_estimator,
+                          feature_names=feature_names,
+                          **kwargs)
+    out.estimator = repr(estimator)
+    return out

eli5/sklearn/transform.py

@@ -0,0 +1,30 @@
+"""transform_feature_names implementations for scikit-learn transformers
+"""
+
+import numpy as np  # type: ignore
+from sklearn.pipeline import Pipeline  # type: ignore
+from sklearn.feature_selection.base import SelectorMixin  # type: ignore
+
+from eli5.transform import transform_feature_names
+from eli5.sklearn.utils import get_feature_names as _get_feature_names
+
+
+# Feature selection:
+
+@transform_feature_names.register(SelectorMixin)
+def _select_names(est, in_names=None):
+    mask = est.get_support(indices=False)
+    in_names = _get_feature_names(est, feature_names=in_names,
+                                  num_features=len(mask))
+    return [in_names[i] for i in np.flatnonzero(mask)]
+
+
+# Pipelines
+
+@transform_feature_names.register(Pipeline)
+def _pipeline_names(est, in_names=None):
+    names = in_names
+    for name, trans in est.steps:
+        if trans is not None:
+            names = transform_feature_names(trans, names)
+    return names

eli5/transform.py

@@ -0,0 +1,29 @@
+"""Handling transformation pipelines in explanations"""
+
+from singledispatch import singledispatch
+
+
+@singledispatch
+def transform_feature_names(transformer, in_names=None):
+    """Get feature names for transformer output as a function of input names
+
+    Used by :func:`explain_weights` when applied to a scikit-learn Pipeline,
+    this ``singledispatch`` should be registered with custom name
+    transformations for each class of transformer.
+
+    Parameters
+    ----------
+    transform : scikit-learn-compatible transformer
+    in_names : list of str, optional
+        Names for features input to transformer.transform().
+        If not provided, the implementation may generate default feature names
+        if the number of input features is known.
+
+    Returns
+    -------
+    feature_names : list of str
+    """
+    if hasattr(transformer, 'get_feature_names'):
+        return transformer.get_feature_names()
+    raise NotImplementedError('transform_feature_names not available for '
+                              '{}'.format(transformer))

tests/test_sklearn_explain_weights.py

@@ -50,7 +50,9 @@
     AdaBoostRegressor,
 )
 from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
-from sklearn.base import BaseEstimator
+from sklearn.base import BaseEstimator, clone
+from sklearn.pipeline import make_pipeline
+from sklearn.feature_selection import SelectKBest
 from sklearn.multiclass import OneVsRestClassifier
 import pytest
 
@@ -484,3 +486,30 @@ def test_feature_importances_no_remaining(clf):
     for expl in format_as_all(res, clf):
         assert 'more features' not in expl and 'more …' not in expl
         assert 'x1' not in expl  # it has zero importance
+
+
+@pytest.mark.parametrize(['transformer', 'X', 'feature_names',
+                          'explain_kwargs'], [
+    [None, [[1, 0], [0, 1]], ['hello', 'world'], {}],
+    [None, [[1, 0], [0, 1]], None,
+     {'vec': CountVectorizer().fit(['hello', 'world'])}],
+    [CountVectorizer(), ['hello', 'world'], None, {'top': 1}],
+    [CountVectorizer(), ['hello', 'world'], None, {'top': 2}],
+    [make_pipeline(CountVectorizer(),
+                   SelectKBest(lambda X, y: np.array([3, 2, 1]), k=2)),
+     ['hello', 'world zzzignored'], None, {}],
+])
+@pytest.mark.parametrize(['predictor'], [
+    [LogisticRegression()],
+    [LinearSVR()],
+])
+def test_explain_pipeline(predictor, transformer, X, feature_names,
+                          explain_kwargs):
+    y = [1, 0]
+    expected = explain_weights(clone(predictor).fit([[1, 0], [0, 1]], y),
+                               feature_names=['hello', 'world'],
+                               **explain_kwargs)
+    pipe = make_pipeline(transformer, clone(predictor)).fit(X, y)
+    actual = explain_weights(pipe, feature_names=feature_names,
+                             **explain_kwargs)
+    assert expected._repr_html_() == actual._repr_html_()