In this fourth (and hopefully final) entry in my “Icing the Kicker” series of posts, I’m going to jump back to the first post where I used
tidymodelsto predict whether or not a kick attempt would be iced. However, this time I see if using the
h2oAutoML feature and the
SuperLearnerpackage can improve the predictive performance of my initial model.
The results are just about what I would have expected: they provide a good floor but a human with knowledge of the data and skill with techniques can still beat out-of-the-box AutoML processes. Still, knowing what that floor is can help a lot: run some AutoML tool for a few minutes/hours/days and you have an easy way of letting the business side know the expected model quality. If AutoML already exceeds expectations, you’re golden. If AutoML is close to expectations (on either end, just above or just below), you as a skilled human should be able to improve things a bit more, especially once you have a chance to analyze what the AutoML processes did. If AutoML is way below business expectations of quality, perhaps this isn’t the best project to spend time on. H/T R-Bloggers.