I have a dataset with an imbalanced binary target. One class accounts for about 94 % of the target variable. I used SMOTE to oversample the minority class but after the oversampling step when I train a Random Forest on the oversampled data and make predictions on the test set, it predicts the minority class for the whole test set. I can't seem to figure out where things went wrong or what the problem could be. Any help would be appreciated.

  • 1
    $\begingroup$ Oversampling is a completely invalid statistical technique and represents a misunderstanding of proper accuracy scoring rules. See for example this. $\endgroup$ – Frank Harrell Sep 15 '18 at 11:37
  • $\begingroup$ Have you tried another oversampling technique? What is your classifier's response to the training set? Is the final distribution around 50 %? And, are you certain that you trained and validated your classifier well with other hyper-parameters? Might worth to try another classifier btw just to see if you suffer from the same problem . $\endgroup$ – gunes Sep 15 '18 at 11:42
  • $\begingroup$ Why not consider the possibility that the best prediction is indeed whichever class accounts for 94% of the target regardless of your features - something you won't discover if you oversample? $\endgroup$ – jbowman Sep 16 '18 at 0:54

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.