My theory is that there were outliers in the private test, but @VIRADUS post and all the comments made Zindi realize that the host would end up with a useless model.

If you look at Amy's response to that post, he said, '...we will reveal that the private leaderboard will show a distribution that will be useful to the client, where potential outliers are taken into consideration...'

This statement made it clear to me that they would remove the outliers from the private test.

Yeah, great comp. The outliers in the public LB confused most people. This is my first time seeing outliers intentionally injected into the test set (public). But again, really interesting, and models a real world scenario pretty well.

#keeplearning

The boosting method was the best; I tested my CV; bagging was so much better, but when I saw seen private score, my first benchmark catboost was the best (105) with only num_cols. I don't know how to test and train data correlations. useless effort!

For me, I was too focused on how to manage outliers at the pre-processing level that I forgot that the main thing is to have a solid model that manages to generalize across the entire dataset. An error of judgment on my part especially since Zindi tells us "where potential outliers are taken into consideration..."

So disapointed knowing that my first submissions with simple boosting model which generalize well lead you to a private score of 125.

But keep learning from this competition. Great experience.

Everyone in this post: please find the updated discussion with new findings. Feel free to comment @yanteixeira what do you think about the updated findings?

This is disturbing! to be fair for host and participants, they should have changed the datasets yea, but communicate to us