Timeline for Can I use predictions from LOESS as a form of predictor transformation in logistic regression?
Current License: CC BY-SA 4.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 9, 2021 at 14:14 | comment | added | kjetil b halvorsen♦ | If you found this answer helpful, then please consider upvoting and/or accepting it. | |
| Sep 21, 2021 at 1:51 | comment | added | simohayha | Really good advice. Thanks Björn. | |
| Sep 20, 2021 at 12:57 | comment | added | Björn | In the case of a continuous numeric value in your case, the smoothing parameter of the LOESS curve fulfills a similar role. The smoothing parameter (and x in the example above) is chosen based on what value performs well in predicting data not seen during training (i.e. you try different values via cross-validation). | |
| Sep 20, 2021 at 12:56 | comment | added | Björn | For categories one assigns each category a numeric value like this: (overall avg) * x / ( x + (# of records for category)) + (avg. outcome for the category) * (# of records for category) / ( x + (# of records for category)). I.e. it's a weighted avg. of the category avg. & overall avg. (as if there were an extra x observations in the category with value of overall mean). E.g. if you just used the target avg., then if the outcome is just 0 or 1 and you see a record with a category numerically represented as 0 or 1, you would otherwise know the answer for the record (while this way you don't). | |
| Sep 20, 2021 at 12:46 | comment | added | simohayha | Thanks Björn. Could you please provide more info regarding "with various ideas like shrinking towards the mean to help avoid overfitting/target leakage"? As for the suitable splines, maybe I will try to find a spline with a monotonic trend (not sure if there is only one or not). | |
| Sep 20, 2021 at 7:55 | history | answered | Björn | CC BY-SA 4.0 |