Andrei Alexandrescu wrote: [ About ranks] > Actually they need to be computed.
Then the problem is still too unclear. > I think it's possible to find cases where > rank(a, k1) + rank(a, k2) < rank(b, k1) + rank(b, k2) > but > alpha * a.k1 + beta * a.k2 > alpha * b.k1 + beta.k2. Of course. Especially if the a, b, ... are sortable only partielly. But if the ranks are known, the problem can be finished with the linear median algorithm. > One potentially confusing issue is that importance applies to > rank, not features. Do you mean that alpha * rank(a, k1) + beta * rank(a, k2) has to be extremized for all `a'? Again: the problem is still unclear. -manfred