Skip to content

Including Scaling and PCA as parameter of GridSearchCV

I want to run a logistic regression using GridSearchCV, but I want to contrast the performance when Scaling and PCA is used, so I don’t want to use it in all cases.

I basically would like to include PCA and Scaling as “parameters” of the GridSearchCV

I am aware I can make a pipeline like this:

mnl = LogisticRegression(fit_intercept=True, multi_class="multinomial")

pipe = Pipeline([
    ('scale', StandardScaler()),
    ('mnl', mnl)])

params_mnl = {'mnl__solver': ['newton-cg', 'lbfgs', 'liblinear', 'sag', 'saga'],

The thing is that, in this case, the scaling would be applied in all folds, right? Is there a way to make it so it’s “included” in the gridsearch?


I just read this answer and even though it’s similar to what I want, it’s not really it, because in that case the Scaler will be applied to the best estimator out of the GridSearch.

What I want to do is, for example, let’s say

params_mnl = {'mnl__solver': ['newton-cg', 'lbfgs']}

I want to run the regression with Scaler+newton-cg, No Scaler+newton-cg, Scaler+lbfgs, No Scaler+lbfgs.



You can set up the parameters with_mean and with_std of StandardScaler() as False to represent no standerdization. In the GirdSearchCV, the parameter para_grid can be set up as

param_grid = [{'scale__with_mean': [False],
               'scale__with_std': [False],
               'mnl__solver': ['newton-cg', 'lbfgs', 'liblinear', 'sag', 'saga'],
              {'mnl__solver': ['newton-cg', 'lbfgs', 'liblinear', 'sag', 'saga'],

Then the first dict in the list is “No Scaler+mnl” and the second is “Scaler+mnl”


Edit: I think it’s complicated if you also considering turn on/off PCA… Maybe you need to define a customised PCA which derives the original PCA. And then define additional boolean argument which determines whether the PCA should be executed or not…

class MYPCA(PCA):
    def __init__(self, PCA_turn_on, *args):
        self.PCA_turn_on = PCA_turn_on
    def fit(X, y=None):
        if (PCA_turn_on == True):
            return super().fit(X, y=None)

    # same for other methods defined in PCA