When can we say random forest is overfitting?

Question

I've divided my data into 3 sets: train, validate and test

I've trained my data once on train set and tested it on the test set. After tuning random forest on validate set, I've got the following numbers:

I'm using accuracy as the evaluation metric as all classes are balanced

Train Acc: 96%

Test Acc: 93%

RF oob_score_: 91%

As there's a 2% gap between my test accuracy and out-of-bag score, is the model overfitting or not? What is the minimum gap at which I can say the model is starting to overfit?

Answer 1

A sign of overfitting is when your model performs really well on the training set, achieving high accuracy, but performs badly on an independent test set. Under the assumption that you've properly split your training, validation, and test sets, and there's a sufficiently large number of samples to justify not using cross validation, then it is highly unlikely that your model is overfitting, as it still achieves relatively high accuracy on the test set.

For more details about overfitting, you can check https://en.wikipedia.org/wiki/Overfitting