[FIXED] Sklearn custom transformer to extract date features doesn't work in a Pipeline

Issue

The data I have in comes in as a string data type, from that string I want to extract some dates that are relevant at prediction time.

An example would be:

Date	City
2015-07-12	Barcelona
2015-07-13	Brussels

And I want it to be

Day	Month
7	12
7	13

etc.

from sklearn.base import BaseEstimator, TransformerMixin

class DateTransformer(BaseEstimator, TransformerMixin):
    def fit(self, X, y=None):
        return self

    def transform(Self, X, y = None):
        X.date = pd.to_datetime(X.date)
        X["year"] = X.date.dt.year
        X["month"] = X.date.dt.month
        X["day"] = X.date.dt.day
        X["dow"] = X.date.dt.dayofweek
        X["quarter"] = X.date.dt.quarter
        X = X.drop("date", axis=1)
        X = X.astype(str)
        return X

It is used in the following pipeline

naiveBaseline = Pipeline([
    ('dates', DateTransformer()),
    ('onehot', OneHotEncoder()),
    ('regression', RidgeCV())
])

When used like this the Pipeline actually works:

naiveBaseline2.fit(X_train,y_train)
naiveBaseline2.predict(X_test)

But ideally I use the following function to benchmark the performance of multiple models:

def evaluate(model, X, y, cv):
    cv_results = cross_validate(
        model,
        X,
        y,
        cv=cv,
        scoring=["neg_mean_absolute_error", "neg_root_mean_squared_error"],
    )
    mae = -cv_results["test_neg_mean_absolute_error"]
    rmse = -cv_results["test_neg_root_mean_squared_error"]
    print(
        f"Mean Absolute Error:     {mae.mean():.3f} +/- {mae.std():.3f}\n"
        f"Root Mean Squared Error: {rmse.mean():.3f} +/- {rmse.std():.3f}"
    )

If I run it through this evaluate function as such: evaluate(naiveBaseline2, X, y, TimeSeriesSplit())

I simply get NaN's for both metrics. I've been going over it for hours and really can't seem to understand what is going wrong, would any of you know?

Solution

Solved it, the issue was wiht the oneHotEncoder, while using TimeSeriesSplit() and walk forward cross-validation it doesn't get to see all date labels (e.g. 2020 in the first split) so it threw an error. The oneHotEncoder needs the parameter handle_unknown=True.

Answered By - Zestar75