Sure, in a concave down horizontal asymptote kind of way. My impression from things like the Netflix prize is that improving predictive accuracy for recommendations is an extremely high-cost, low-yield endeavor. I remember one day being shocked at the quality of Amazon's recommendations for me (since they'd always been crap), and looking to see how they'd picked such winners. The result? All of those products were ones I'd personally added to my old forgotten amazon wishlist.