• data/features_none_sre.csv: No baseline
• data/features_sub_sre.csv: Subtracted by baseline
• data/features_div_sre.csv: Divided by baseline

• data/features_none_srl.csv: No baseline
• data/features_sub_srl.csv: Subtracted by baseline
• data/features_div_srl.csv: Divided by baseline

For Each one, there is a problem of imalancing of the classes, in particular we have:

• 1/5 subject old card (class 1)
• 5/6 new card (class 0)

This imbalancing can cause a problem like: i I have to detect the 1 I'll always tell 0. So the accuracy will be very high but the classifier will be crap.

So the idea is to:

• Rebalance the dataset
• Use different evaluation rather than the accuracy (like the F1_score)

I tried two alternatives:

• Oversample the class 1 (with the SMOTE algorithm)
• Aggregate the class 0 by subject in order to balance the dataset

Another problem with the oversampling is the combined use of cross-validation. If I oversample the dataset before the cross-validation (the problem is bigger with the copying method, but still present with SMOTE) There is the risk to have the same data in multiple folds. Otherwise, with SMOTE I oversample, then split the data and pretend that value is not fabricated.