Fit Once, Scale Twice

You're building a preprocessing pipeline with a StandardScaler before a classifier and splitting data into train and test sets. Which approach avoids data leakage?

Fit the scaler on the test set only, then apply it to the training set

Fit the scaler on the training set only, then use those fitted parameters to transform both train and test sets

Fit the scaler on the full dataset (train + test combined), then split into train and test

Fit a separate scaler on the test set so its distribution is scaled correctly