failed in rename column name and output pmml

[duanzongran]

I want to do a category declaration, missing value impute, target encoding of the character variables in the data, and then rename them with 'cat_' as the prefix(This is something I must do), which is successful for data preprocessing(I mean I got the expected column names) and modeling, but I can't export it to a pmml file properly.

Here is my code:

import pandas as pd
from sklearn_pandas import DataFrameMapper

from sklearn.impute import SimpleImputer
from sklearn.tree import DecisionTreeClassifier
from sklearn.pipeline import Pipeline
from category_encoders import TargetEncoder

from sklearn2pmml import sklearn2pmml, PMMLPipeline
from sklearn2pmml.decoration import Alias, CategoricalDomain



data = pd.DataFrame({
    'cat1': ['A', 'B', None, 'C','C','B'],
    'cat2': ['X', 'Y', 'X', 'Z','X','Y'],
    'target': [0, 1, 0, 1,1,0]
})

X = data[['cat1', 'cat2']]
y = data['target']


cat_origin_list = ['cat1', 'cat2']

cat_mapper = [
    (
        [col],
            Pipeline([
                ("domain", CategoricalDomain(with_data=False)),
                ("imputer", SimpleImputer(missing_values=None,strategy="constant", fill_value="-999")),
                ("target_encode", TargetEncoder())
            ]),{"alias":f"cat_{col}"}
    )
    for col in cat_origin_list
]


mapper = DataFrameMapper(
    cat_mapper,
    df_out=True
)


pipeline = PMMLPipeline([
    ("mapper", mapper),
    ("classifier", DecisionTreeClassifier())
])


pipeline.fit(X, y)
X_transformed=pipeline.named_steps['mapper'].transform(X)

sklearn2pmml(pipeline, "model-test.pmml", pmml_schema="4.3")

It raised the error :

Standard output is empty
Standard error:
Exception in thread "main" org.jpmml.sklearn.SkLearnException: User input field cat1 cannot be renamed
	at org.jpmml.sklearn.SkLearnEncoder.renameFeature(SkLearnEncoder.java:284)
	at sklearn_pandas.DataFrameMapper.encodeFeatures(DataFrameMapper.java:79)
	at sklearn_pandas.DataFrameMapper.initializeFeatures(DataFrameMapper.java:48)
	at sklearn.Initializer.encode(Initializer.java:59)
	at sklearn.Composite.encodeFeatures(Composite.java:112)
	at sklearn.Composite.initFeatures(Composite.java:255)
	at sklearn2pmml.pipeline.PMMLPipeline.encodePMML(PMMLPipeline.java:113)
	at com.sklearn2pmml.Main.run(Main.java:100)
	at com.sklearn2pmml.Main.main(Main.java:85)

[duanzongran]

Thank you for editing the issue @vruusmann .
What I need to add here about the code is : when I delete ',{"alias":f"cat_{col}"}' from cat_mapper ,everything seems OK except the
the need to rename variables

[vruusmann]

I can see that you've tried two ways of renaming a feature:

• Using the "alias" mechanism of the DataFrameMapper meta-transformer.
• Using the sklearn2pmml.decoration.Alias decorator (you have the import, but the actual usage looks to be erased from the sample code).

These two approaches fail with the same exception, stating that the "user input field X cannot be renamed". I analyzed it a bit, but couldn't get it sorted out immediately.

In brief, the TargetEncoder is supposed to create a new DerivedField element based on the DataField element. However, for some reason, the TargetEncoder does not materialize its soon-to-be-generated DerivedField element right away, so the renaming still sees the user input field in that spot, and makes a complaint.

There is a third way of renaming features - simply set a <Transformer>.pmml_name_ (note the trailing underscore character!) attribute.

In your case:

def make_target_encoder(col):
    encoder = TargetEncoder()
    # THIS!
    encoder.pmml_name_ = f"cat_{col}"
    return encoder

cat_mapper = [
    (
        [col],
            Pipeline([
                ("domain", CategoricalDomain(with_data=False)),
                ("imputer", SimpleImputer(missing_values=None,strategy="constant", fill_value="-999")),
                # THIS!
                ("target_encode", make_target_encoder(col)),
            ])
            # Not this - doesn't work correctly with TargetEncoder
            #]),{"alias":f"cat_{col}"}
    )
    for col in cat_origin_list
]

[duanzongran]

I can see that you've tried two ways of renaming a feature:

• Using the "alias" mechanism of the DataFrameMapper meta-transformer.
• Using the sklearn2pmml.decoration.Alias decorator (you have the import, but the actual usage looks to be erased from the sample code).

These two approaches fail with the same exception, stating that the "user input field X cannot be renamed". I analyzed it a bit, but couldn't get it sorted out immediately.

In brief, the TargetEncoder is supposed to create a new DerivedField element based on the DataField element. However, for some reason, the TargetEncoder does not materialize its soon-to-be-generated DerivedField element right away, so the renaming still sees the user input field in that spot, and makes a complaint.

There is a third way of renaming features - simply set a <Transformer>.pmml_name_ (note the trailing underscore character!) attribute.

In your case:

def make_target_encoder(col):
encoder = TargetEncoder()
# THIS!
encoder.pmml_name_ = f"cat_{col}"
return encoder

cat_mapper = [
(
[col],
Pipeline([
("domain", CategoricalDomain(with_data=False)),
("imputer", SimpleImputer(missing_values=None,strategy="constant", fill_value="-999")),
# THIS!
("target_encode", make_target_encoder(col)),
])
# Not this - doesn't work correctly with TargetEncoder
#]),{"alias":f"cat_{col}"}
)
for col in cat_origin_list
]

Your reply solved my problem perfectly! Thank you very much for your patience and kindness.
Also, I found another solution: use sklearn's TargetEncoder instead of using category_encoders.

from sklearn.preprocessing import TargetEncoder

cat_mapper = [
    (
        [col],
            Pipeline([
                ("domain", CategoricalDomain(with_data=False)),
                ("imputer", SimpleImputer(missing_values=None,strategy="constant", fill_value="-999")),
                ("target_encode", TargetEncoder(cv=3, smooth=1.0, random_state=42))
            ]),{"alias":f"cat_{col}"}
    )
    for col in cat_origin_list
]

[vruusmann]

Also, I found another solution: use sklearn's TargetEncoder instead of using category_encoders.

Scikit-Learn's TargetEncoder materlaizes the DerivedField element immediately, so the rename action sees intended field. However, Category-Encoders' TargetEncoder delays the materialization for a while, which then causes problems.

Leaving the issue open until I have more time to investigate this matter more deeply.

Anyway, setting the <Transformer>.pmml_name_ is the most robust approach (although undocumented), because it sets the field name independently of its materialization status.