How to scale train, validation and test sets properly using StandardScaler?

Question

Some articles says that in case of having only train and test sets, first, we need to use fit_transform() to scale training set and then only transform() for test set, in order to prevent data leakage.

In my case, I have also validation set.

I think one of these codes below would be okay to use but I cannot rely on them completely. Any kind of help will be appreciated, thanks!

1)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.3)
X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size = 2/7)
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_val = scaler.transform(X_val)
X_test = scaler.transform(X_test)

2)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.3)

scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)

X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size = 2/7)
X_test = scaler.transform(X_test)

Calfskin

First code is generally considered best practice. Fitting the scaler on only the training data prevents data leakage between your model training and model validation – Oram

Number 1) is correct. In terms of scaling you should treat test and val data in the same way. – Linetta

Thank you for your help! I got the point without any question mark. – Calfskin

Answer 1

Generally you would want to use Option 1 code. The reason for using fit and then transform with train data is a) Fit would calculate mean,var etc of train set and then try to fit the model to data b) post which transform is going to convert data as per the fitted model.

If you use fit again with test set this is going to add bias to your model.

Toxoplasmosis

Answer 2

Scale dataset : logistic regression, support vector machine and random forest . How are the results compared to the lupin ski descriptors

Inopportune