I think that at the core of this problem is: 1.) Separate evaluation is done in absolute terms. 2.) Join evaluation is done in relative terms in a specific mo

I think that at the core of this problem is: 1.) Separate evaluation is done in absolute terms. 2.) Join evaluation is done in relative terms in a specific model. Which specific model? The model that only takes into account the union of features of both items. So this is similar to a relative comparison with conditional probability of occurrence. The relative evaluation is done thinking about: In which situations would item B be more valuable than item A. All those situations are weighted according to the avaiability heuristic: easy to recall situations seem more frequent or important. To give a simple example: Perhaps a dictionary with 20000 words is enough for you, and you don't need a greater dictionary. But you can think than may words are in the 40000 words dictionary but not in the 20000 word dictionary, so the difference seems to be more important that what really is for you.