At what stage we can confirm the model is overfitting?

Module 2. Selecting the best model Module overview
#1

Hi Team,

My question is around overfitting, let’s say I have trained the model and evaluating it with the test how can identify whether the model is overfitting? I know the overfitting concept, but I want to understand what difference of accuracy of the model between the train and the test says the model is overfitting?

Eg: 1. the training set produced 92 % of accuracy and the test produced 85% of accuracy and difference is 7%. Do we say this model is overfitting? is there any standard to identify the model is overfitting with the variance ± value?

Eg: 2. the training set produced 85 % of accuracy and the test produced 92% of accuracy. Do we say this model is overfitting?

#2

When the testing score is worse than the training score then we say that the model overfit. Then, you can have a model that slightly overfits but that still perform well.

In practice, I don’t think that there is such a case that you better generalize on the testing set than on the training. The only way to have these results would be that the training dataset contains some noise and we are able to regularize the model to avoid fitting the noise and that the testing contains less noise than the training set.

#3

thanks very much for your reply :slight_smile: