I use a couple interaction terms in some of my ranking systems using FRank(A)*FRank(B) to magnify the effect when both factors are high (or both are low). There's a little more color in this thread, and in one of Marco's replies, he highlights why this approach is different from a weighted average inside of a Composite:
Jim, I'll have to give your theory on whether a top 15 optimization would capture this if it was based on a weighted ranking of factors. My intuition says no as that relationship couldn't be represented in the linear formula of individual terms but I have to give it more thought.