Timeline for Target encoding in test data and target leakage
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 9, 2022 at 14:31 | comment | added | Ben Reiniger | Oh! unless in your quote the outer 20-fold cv is the model evaluation cv; I had been interpreting the whole discussion as happening in the training set only, with a separate testing set. Then I think the rest of my answer actually agrees with that source. | |
| Mar 9, 2022 at 13:46 | comment | added | Ben Reiniger | @Chris (1) I just mean the average target for the category (with smoothing if you used that), though that is equivalent (the average of leave-one-out averages is just the flat average). (2) maybe? I'll think some more about it, but just the outer cv already reduced that overfitting potential. (3) I meant what Björn described in the regularization section. | |
| Mar 9, 2022 at 13:34 | comment | added | Chris |
Thanks for the link! I'm already reading it!. (1) When you say "usually you just use the average for the training set" you mean the average of the target encoding per category value? Specifically, for B in the @bjorn example, would be (0.67 + 0.67 + 0.33 + 0.33)/4=0.5 the value for B in the test set? (2) Doesn't that nested cross validation reduce the overfitting that you say it isn't necesary? (3) Could you point me to examples about "smoothing with the prior"? I find many different "smoothing with the prior" in google and I don't know which one you are refering to..
|
|
| Mar 8, 2022 at 18:43 | history | edited | Ben Reiniger | CC BY-SA 4.0 |
added 3 characters in body
|
| Mar 8, 2022 at 18:29 | history | answered | Ben Reiniger | CC BY-SA 4.0 |