I am facing a two-class classification problem where:

  1. There is way more training data in class 1 than in class 0.

  2. Classifying a class 0 event as class 1 has a higher loss than classifying a class 1 event as class 0.

  3. The prior probability of an event belong to class 0 is about 50 times smaller than the probability of belonging to class 1.

The way I see it I have (at least) 4 mechanisms to address these issues:

a. Collecting more data for class 0 than class 1 makes the classifier biased towards class 0.

b. Weighting the samples in class 0 higher than the ones in class 1.

c. Introducing an asymmetric loss function (which I understand as being identical to a 0-1 loss function in the weighted class case)

d. Given that the classifier gives returns a probability with which it assumes that a sample belongs to class 0, use a threshold smaller than 0.5 to assign a sample to class 0.

To me it seems like all four mechanisms are capable of addressing issues 1) and 3) (but not 2)) so I am wondering what is the cleanest way of addressing this?


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.