there doesn't seem to be a ton of improvement in a given model across different runs of RandomizedSearchCV. It might improve by a couple single-digit percentage points, but not dramatically.
I think the most valuable thing we can do by default is to surface which models are useful, and which ones just aren't. And then, of course, to let ensembling take over.
Single-digit percentage point increases in accuracy are huge in some circumstances, but at that point, we'll let the engineers go in and focus that level of improvement on only the models that matter to them. It's an easy adjustment to make. Maybe write a blog post about this.
there doesn't seem to be a ton of improvement in a given model across different runs of RandomizedSearchCV. It might improve by a couple single-digit percentage points, but not dramatically.
I think the most valuable thing we can do by default is to surface which models are useful, and which ones just aren't. And then, of course, to let ensembling take over.
Single-digit percentage point increases in accuracy are huge in some circumstances, but at that point, we'll let the engineers go in and focus that level of improvement on only the models that matter to them. It's an easy adjustment to make. Maybe write a blog post about this.