| advertisement: compare things at compare-stuff.com! |
One work-around to the problem of architectures with few representatives is to perform a two-way prediction between the 17 ribbon and 112 non-ribbon architectures. The Matthews correlation coefficient for this prediction is 0.614, which is worse than the previous multi-architecture prediction. In terms of composition centroid location, the 112 domains of different architectures may be closer to the ribbon domains than each architecture alone. The predictions are 91% accurate overall (78% random accuracy), but with only 71% of the ribbon domains correctly predicted. The top two quartiles are both 100% accurate, and only one of these predictions is for a ribbon domain. Thus we may be able to predict non-ribbon confidently on many occasions, but may remain undecided about many others which may have ribbon architectures.