I wrote a note on how to combine multiple predictors with only marginally higher than baseline accuracy to construct an ensemble predictor that is much more accurate, via unanimous voting. It has a few simple formulas which don’t work in html so here it is in pdf format.
I’m a bit worried because it looks much more mathematically complicated than it is (it’s only simple fraction arithmetic and the most basic probability) and because it has no pictures. There is a table at least. It’s written to be as intuitive as possible though, like everything else.
I’d like to hear if anyone’s using techniques like this.
