This is more of a hypothetical than something I'm actively trying to solve. It just struck me that a machine learning algorithm that specifically looked at two pieces of data and had to label one as greater than the other or something of that nature might be inherently different than classifying each separately and comparing the strength of the classifications.
There are ranking algorithms based on machine learning that are aimed to build ranking models. Training data for these models is given in the form of partial ordering between each pair of elements in a sample. A brief description, together with a list of useful references, is given in the corresponding Wikipedia page.
Not sure if I read the question correctly, but I have implemented multi-variate sub-sample optimisation routines using the Kolmogrov-Smirnov test?
The advantage is that if two distributions exactly track each other (so would be significantly correlated) but are at quite different scales (e.g. one is half the value of the other), they would still rank as very different distributions at similar scales. Only when both distribution and scale match is the ranking high.
Spatially per-pixel comparison algorithms are also quite common in raster analysis?