what is your offline score about kaggle-quora-dup

[aerdem4]

Hi, thanks for your question. The validation error you see is on one fold. When you run it for all the folds and take average of the models generated in each fold you get some gain from ensembling. In addition, if you look at postprocessing, I do some extra things to improve my score. As a last remark, logloss is a class balance sensitive metric, therefore logloss on train/val set and test set (Kaggle LB) can be very different since Kaggle didn't split them randomly.

[sanshanxiashi]

Thank you very much for your reply!
I understand what you mean. By the way, how does the correct rate of the test set be estimated? Is the probability greater than 0.9?
@aerdem4

[aerdem4]

I think they probed it by submitting all 1s or all 0s.