-
|
Hi @rasbt , I achieve it. But what bothers me is that validation accuracy is greater than training accuracy. How's it even possible? Intuitively training accuracy should always be higher than the validation (unseen data) accuracy right? Could you please shed some light on this?
I used |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
That's an interesting one. It's rare but it can happen. This could be due to randomness (Ultimately, it's still a relatively small dataset, and there are only ~270 instances in the validation set). E.g., assume the model does not overfit and performs equal on the training and validation set. You will always have some slight variation based on which exact examples are in the training and the validation (i.e., how you split the dataset by varying the random seed, for example). Another explanation is that you have tuned the model so much (via hyperparameter choices) to improve the validation accuracy that it is slightly overtuned on the validation set. This is why it's important to have an additional test set for the final evaluation (If I recall correctly, I covered it later in the course.) |
Beta Was this translation helpful? Give feedback.
-
|
Nice explanation. Both of the points make sense. |
Beta Was this translation helpful? Give feedback.
That's an interesting one. It's rare but it can happen. This could be due to randomness (Ultimately, it's still a relatively small dataset, and there are only ~270 instances in the validation set). E.g., assume the model does not overfit and performs equal on the training and validation set. You will always have some slight variation based on which exact examples are in the training and the validation (i.e., how you split the dataset by varying the random seed, for example).
Another explanation is that you have tuned the model so much (via hyperparameter choices) to improve the validation accuracy that it is slightly overtuned on the validation set. This is why it's important to have an a…