On previous episodes I said:

A bit further along the road, I tried an ensemble of classifiers, which is coded in classifier_ensemble.R, but I could not get anything useful from there.

Ok, not really. I wasn't seeing any improvement on the public leaderboard, but when the competition ended I checked the private score for those submissions. Oh, surprise surprise:


The third column is the score for the public leaderboard; the fourth column is the score for the private leaderboard (ultimately, the one used when the competition ends.) Turns out I was getting an excellent score! Had I selected the top submission I would have landed around the 17th position on the final leaderboard.

So, yes, model ensembling was working just fine. It was just that I could not see it.