I built a binary classification model. I split data into train and test and I scaled them (StandardScaler, fit_transform on train - transform on test). Everything was aimed at achieving a good scoring (recall in my case). After a GridSeachCV, I ended up the best model was XGBoost with the optimal hyperparameters identified by GridSearchCV.

So my questions are more on the process:

  • Should I retrain that XGBoost model on the entire dataset (model.fit(X_data, y_data))? I guess yes.
  • Should I use cross-validation at this stage again? if yes, why? and how?
  • I have to scale my data as I did it during the model construction, but should I recalculate the scaling based on the full dataset or should I use the scaling I computed during the model construction?


  • $\begingroup$ Recall is synonymous with sensitivity, isn’t it? If all you want to do us maximize sensitivity, why not make your life really easy by skipping all of this, always predicting the class you’re trying to find, achieve perfect recall, and charge an exorbitant consulting rate as the Master of Recall? $\endgroup$ – Dave Feb 17 at 14:37
  • $\begingroup$ ehehehe good point.Yes recall= sensitivity or true postive rate. Anyway, recall is just an example. It is a little bit more complex of how I described it, but my question is on the process after the parameter tuning, whatever the scoring $\endgroup$ – Luigi87 Feb 17 at 14:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.