Timeline for Does oversampling/undersampling change the distribution of the data?
Current License: CC BY-SA 4.0
14 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 19, 2019 at 14:12 | answer | added | Christopher John | timeline score: 1 | |
| Oct 18, 2019 at 5:32 | answer | added | etudiant | timeline score: 1 | |
| Sep 25, 2019 at 17:12 | answer | added | Ben Reiniger | timeline score: 8 | |
| Sep 25, 2019 at 14:53 | comment | added | Anuj | @mkt I appreciate the help! Thanks a lot! | |
| Sep 25, 2019 at 14:48 | comment | added | mkt | I posted the links because improper scoring rules are a common topic of discussion here. The second link focuses on accuracy but it illustrates some of the problems with improper scoring rules in general. It's not my area, so I'll leave it at that. | |
| Sep 25, 2019 at 14:40 | comment | added | Anuj | @mkt thank you for the links but I don't see how they answer my question. The first link has a post about how f-1 score is not an ideal performance metric for imbalanced classification. But I have read numerous papers, like this which prove that f-1 score is a good performance metric. The second link has good points about why accuracy is not the best measure, which I am well aware of, and that is exactly why I am not using accuracy to measure the performance of the model. | |
| Sep 25, 2019 at 14:21 | comment | added | mkt | My mistake. See stats.stackexchange.com/a/210718/121522 and stats.stackexchange.com/questions/312780/… | |
| Sep 25, 2019 at 14:17 | comment | added | Anuj | @mkt It is 0.83 for the positive class which is the majority but only 0.13 for the negative (minority) class. It classifies data from the positive class as the negative class. | |
| Sep 25, 2019 at 13:55 | comment | added | mkt | Why do you think 0.83 is poor? It's normal for performance on the test set to be worse than on the training set | |
| Sep 25, 2019 at 13:55 | comment | added | mkt | stats.stackexchange.com/questions/283170/… | |
| Sep 25, 2019 at 13:46 | comment | added | Anuj | Yes, this was my understanding too. But why is the model performing so poorly on the test set even though it has learnt features from the train set? Also, if I perform oversampling/undersampling on the test set as well, I get good results on both the positive and negative classes. This led me to believe that sampling changes the distribution of the data. | |
| Sep 25, 2019 at 13:37 | comment | added | user2974951 | Over/undersampling doesn't add any new information, it only replicates data, which is done to prevent the model from being biased, but still doesn't help the model to learn better. | |
| Sep 25, 2019 at 13:35 | review | First posts | |||
| Sep 25, 2019 at 14:35 | |||||
| Sep 25, 2019 at 13:31 | history | asked | Anuj | CC BY-SA 4.0 |