Let's say I am a software engineer, and I implemented a random forest classifier.

Now, I have trained the classifier and I want to make prediction on an observation. The majority votes are supposed to be based on OOB samples. Instead, I used all samples.

In math, what's the negative effects? Does this really make a difference in real world?


If you predict a new observation, you send it through all trees and then bundle the results by majority votes. You seem to do it like this, which is good.

OOB predictions are relevant to guess the model performance on the training data. This is of key importance since random forests usually grossly overfit on the training data. Thus, missclassification rates would be much too good. If your implementation cannot do this, you can estimate the model performance through cross-validation.


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.