Imagine the following scenario.

Train a classifier that classifies an object into one of these n+m classes:

class 1..n     -> triggers an action A depending on the predicted class p
class n+1..n+m -> triggers an action X independent on the prediction

Used is the CalibratedClassifierCV of a LinearSVC of scikit-learn. So, the prediction is based on the maximum predict_proba value.


Because action X is independent on the prediction, I would very like to combine the m classes (n+1, n+2, ..., n+m) into a single class (to improve training speed, classification speed, memory consumption etc.)

Would that harm the classification result?


So what you are interested in (in terms of prediction) is $n+1$ classes, where the last class is $m=1$. This would be my default model. In other words, I don‘t see why you should have $m$ additional outcomes if these outcomes are irrelevant for your task.

For the quality of a prediction, features $X$ matter the most. I‘m not aware that the classes are relevant unlike your outcome (=classes) have an order.

However, ultimately you need to test which model performs best. So by the parsimony principle, start with a simple model and gradually see if adding complexity delivers a better result (by a predefined measure such as accuracy).


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.