Timeline for How to make a classification problem into a regression problem?
Current License: CC BY-SA 4.0
4 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 27, 2020 at 13:22 | comment | added | Ben Reiniger♦ | Marijn, One-hot encoding the target but keeping it classification is just using the one-vs-all multiclass approach. Going from there to regression on 0/1 variables is not advisable: e.g., a score of -0.231 will get penalized for being "away" from zero the same amount as +0.231. @EdoardoGuerriero, it isn't the same as modeling logits though. | |
| Mar 27, 2020 at 12:41 | comment | added | Edoardo Guerriero | By doing this you're not turning a classification problem into a regression one, you're just stopping to the logits instead of returning probabilities, i.e. if you apply the softmax and then argmax to the scores you wrote you arrive at the same point where you started. One hot encode does not change the labels making them suitable for regression, to do so you need to turn discrete values into continuous one in the initial dataset, that's the defenition of regression. | |
| Mar 27, 2020 at 12:05 | review | First posts | |||
| Mar 27, 2020 at 12:09 | |||||
| Mar 27, 2020 at 12:00 | history | answered | Marijn van Vliet | CC BY-SA 4.0 |