bloods.ai Blood Spectroscopy Classification Challenge
Helping Global
$7 500 USD
Completed (~4 years ago)
Classification
1103 joined
265 active
Start
Oct 19, 21
Close
Feb 13, 22
Reveal
Feb 13, 22
Congratulations
Platform · 14 Feb 2022, 06:50 · 7
Dear all,
Congrats to all for finishing this challenging competition (I personally found this competition really really hard). Many thanks to my teammates @Skaak and @Azer.
BTW, it is so weird that some top teams have ZERO submissions :-) Probably people submitted and got a score before the submission counter was reset @Zindi should really re-run all the provided code and re-rank the LB based on the "new" scores.
Hi Moto,
We will look into this.
Congrats!
Could You describe the solution (or share the code)?
As for me competition's data seems to have no signal at all (models don't understand how to separate different classes). I tried different preprocessing settings, tree models, DL models, PCA/SVD/RBM features but with no effect.
I guess You did something special here.
You know - we have saying here in Africa. The elephant and the mosquito cross a bridge - boy do they shake it. In this competition, @Moto is the elephant and this time there was an @ASSAZZIN riding him. I was mosquito who had privilege to test and verify there work and learned so much. As mentioned, we used very simple (or very aggregated) approaches and this outperformed others, except for some really beautiful nonlinear stuff that @ASSAZZIN was able to concoct and that gave us nice improvement toward the end. Boy did we shake that bridge.
Congrats!
Thanks you @Amy. It is also great to make sure all the code does not have any unnormal piece.
@NikitaChurkin: Thanks for interesting in our solution. @Skaak, our team brain, is working on it.
@NikitaChurkin - as @Moto says, this was really really hard. I think there is distro shift between train and test, and thought you might win given excellent Kolmogorov-Smirnoff optimisation you did on kaggle for similar shift in brain image competition. The way we handled this was using simple approach to prevent overfit, selecting low standard deviation channels (based on some papers about blood spectroscopy you can find online) and then also ensembling.
It was an amazing competition ! I hope we get some insights about users' behavior, especially after this one !