0
$\begingroup$

this might be very beginner's question. I'm working on Kaggle's HomeCredit Default Risk problem which has among others dataset train, test and submission files as can be seen in the link provided. The test dataset does not contain TARGET feature and submission has ID and TARGET columns where TARGET has fixed value 0.5.

I'd normaly split the train dataset and used it to train and score a model and than created a submission file based on the test dataset. Thus, I don't understand what I am supposed to do with the already existing submission file.

Futher, when I score a model based on splitted train data, I'm getting unexpected high values which can't be right. I did before OHE-encoding, standardization and PCA, it that might have some impact on the model accuracy result.

Insights would be apprecited.

$\endgroup$
0
$\begingroup$

The submission file is just there as a reference. You are not supposed to do anything with it, but when you make your own submission it be on the same format.

Regarding your high accuracy. Try making your own submission file and submit it to Kaggle and it will show you if your score is real or not. :)

link for submission

$\endgroup$
  • $\begingroup$ There was class imbalance problem. AUROC score is way more realistic. $\endgroup$ – Stanislav Jirak Mar 18 '19 at 16:06
  • $\begingroup$ Yes, accuracy can be misleading. :) Happy kaggling! $\endgroup$ – Simon Larsson Mar 18 '19 at 18:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.