Timeline for Machine learning - 'train_test_split' function in scikit-learn: should I repeat it several times?
Current License: CC BY-SA 4.0
8 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 22, 2018 at 13:40 | comment | added | qmeeus | Indeed, but you cannot be lucky or unlucky all the time... As I told you, never witnessed that in the past. Generally, two runs always give me similar results. If you want to check it out, it's easy: just run your code twice and check whether the results are very different. | |
| Aug 22, 2018 at 13:35 | comment | added | Marcello P. | @id-2205: but random is random... if the split is "unlucky", the result will be worse than with a "lucky" split... or not? Where am I wrong? Thank you! | |
| Aug 22, 2018 at 13:22 | comment | added | qmeeus | It does not depend on the split if the split is random... | |
| Aug 22, 2018 at 12:20 | comment | added | Marcello P. | Yes, but... I can't understand how many times I should repeat the procedure composed of steps 1-4 (in the id-2205's answer): is the result of step 4 (Evaluate on the test set) the final result of my analysis? Doesn't it strongly depend on the split made at step 1? Really many thanks | |
| Aug 22, 2018 at 12:08 | history | edited | qmeeus | CC BY-SA 4.0 |
added 348 characters in body
|
| Aug 22, 2018 at 12:03 | comment | added | qmeeus | You can change the seed using the argument 'random_state' of train_test_split... | |
| Aug 22, 2018 at 12:02 | comment | added | Mephy | @MarcelloP. you don't have to change the seed manually, it's there to allow reproducibility. If you're just training a model to be used in production, let it be random. | |
| Aug 22, 2018 at 11:48 | history | answered | qmeeus | CC BY-SA 4.0 |