Though it's clearly questionable from the get-go that an algorithm like Keegan can accurately determine which images are
good (defining and measuring "good" being the key problem), it seems more feasible that it might be able to predict which ones will be
popular.
As a limited, non-rigorous attempt to test that idea, I went back through several several months of images posted in the LuLa Critiques and Landscape & Nature forums. I fed about 60 images to Keegan that were popular in those forums, as indicated by larger than average numbers of positive comments.
I could see no consistent correlation between Keegan's ratings and the apparent ratings conferred by LuLa reviewers.
But I did notice that Keegan does not seem to favor complexity, low contrast, subtlety, muted colors--or "street" photos of almost any sort. For many abstracts, it strangely complains that it can't analyze them at all because they are not photographs. On the other hand, it does seem to be rather fond of well-lit, high contrast, boldly colored, sharp, simple and large main subjects.
Take home lesson: it is easier to divine the image preferences of Keegan than those of LuLa reviewers.