1
$\begingroup$

I've come up with some simple definitions for training, testing and validation data in supervised learning. Can anyone verify/improve upon my answers?

Training Data - Used by the model to learn parameters and 'fit' to the data (usually involves multiple models fit at once)
Validation Data - Used by the model to either a) determine the best hyperparameter(s) for a given model or b) determine the best performing model out of a given selection or c) determine the best hyperparameters of the best performing model out of a given selection (combining a and b)
Testing Data - Used to gauge the final model performance (No further fitting is allowed at this point, so that we have an objective assessment of the model we've trained)

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ Overall correct. About validation data there is not much of "or" and "either". It is actually the third option which is model selection AND hyper parameter tuning. In the "Training" definition, I do not why it "usually involves multiple model to fit at once"! Do not bias yourself with that. $\endgroup$
    – Kasra Manshaei
    Commented Mar 7, 2022 at 9:00
  • 1
    $\begingroup$ @KasraManshaei's comment is sound. In addition, validation data falls under the umbrella of regularization techniques, which, roughly speaking, prevents your model to overfit data. $\endgroup$
    – Eduard
    Commented Mar 7, 2022 at 14:07

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.