Slep007 - feature names, their generation and the API

slep007/proposal.rst

@@ -0,0 +1,57 @@
+ .. _slep_007:
+
+=============
+Feature Names
+=============
+
+`scikit-learn/#13307 <https://github.com/scikit-learn/scikit-learn/pull/13307>`_
+proposes a solution to support and propagate feature names through a pipeline.
+
+However, the generated feature names can become long. This is to confirm how
+we want those generated feature names to behave, and this is the proposal::
+
+
+    from sklearn.pipeline import make_pipeline
+    from sklearn.preprocessing import StandardScaler
+    from sklearn.decomposition import PCA
+    from sklearn.compose import make_column_transformer
+    from sklearn.datasets import load_iris
+    from sklearn.linear_model import LogisticRegression
+    from sklearn.feature_selection import SelectKBest
+    import pandas as pd
+
+    iris = load_iris()
+    X = pd.DataFrame(iris.data, columns=iris.feature_names)
+    pipe = make_pipeline(StandardScaler(), PCA(), SelectKBest(k=2),
+                         LogisticRegression())
+    pipe.fit(X, iris.target)
+    pipe[-1].input_features_
+    > array(['pca0', 'pca1'], dtype='<U4')
+
+
+    # I have to duplicate StandardScaler if I want to be able to
+    # use the pandas column names in the column transformer. Yuk!
+    # also there's no easy way to pass through the original columns with
+    # ColumnTransformer - FunctionTransformer would remove feature names!
+    pipe = make_pipeline(make_column_transformer((make_pipeline(StandardScaler(),
+                                                                PCA()), X.columns),
+                                                 (StandardScaler(),
+                                                 X.columns[:2])),
+                         SelectKBest(k=2), LogisticRegression())
+    pipe.fit(X, iris.target)
+    pipe[-1].input_features_
+    ```
+    > array(['pipeline__pca0', 'standardscaler__sepal length (cm)'], dtype='<U33')
+
+    pipe = make_pipeline(make_column_transformer((PCA(), X.columns),
+                                                 (StandardScaler(),
+                                                 X.columns[:2])),
+                         SelectKBest(k=2), LogisticRegression())
+    pipe.fit(X, iris.target)
+    pipe[-1].input_features_
+
+    > array(['pca__pca0', 'standardscaler__sepal length (cm)'], dtype='<U33')
+
+Is that what we want? (apart from changing to object dtype lol)
+The first one seems good to me, the others seem a bit long? Not sure how to do
+better though.