2

I have done 10-Fold CV on my data and I have selected my best model from the result. With cross validation, I will be having 10 models trained with different folds of data. For final model to submit, should I

take the average of these models or just fit a model for entire training set?

william007
  • 897
  • 1
  • 6
  • 9
  • 3
    I suspect that there is a fundamental misunderstanding about cross validation behind this question which may have been answered already at https://stats.stackexchange.com/q/52274 – cbeleites unhappy with SX Jul 21 '19 at 16:53

1 Answers1

1

Your "final" model, at least based on the 10-fold CV that you performed, is supposed to be the "best" model that you found as a result of your CV.

E.g. (I could be more specific if you provided a reproducible example), if you are testing the performance of a certain number of regression models using 10-fold CV, and as a measure of predictive accuracy you are using the Mean Absolute Percentage Error, i.e. MAPE (for instance), then your "best" model is the one producing the lowest MAPE when averaged across all folds (instead of the average you could also consider the median or any other metric of choice to find how each model performs across the folds). Once that model is identified, then as the last step you can fit it to the entire training data, and suggest it as the "final" model based on the data that you have.

davo.biainili
  • 353
  • 1
  • 7