Our approach is based on boosted trees with the particularity of training different models depending on the actual features present in the current test well. The final model is obtained by soft voting of the best models in validation.
The competition was very close and the results of the top 7 teams are almost equivalent. We are very happy with our result in both leaderboards. Taking part in this competition was a lot of fun and very informative. Thanks to Peter Bormann, Peder Aursand and all the organizers. And thanks to all the teams that made the competition so exciting and that have now shared their solutions. We are convinced that the competition has initiated many positive dynamics that will continue for a long time!
See the codes of the top-10 solutions on GitHub!