Amazon recommendations

A regular refrain in the the discussion going on hosted by Charlie Stross on Amazon's e-book strategy is little paeans to the Amazon recommendations; which I find odd, because from my experience Amazon recommendations fail badly, in a number of ways.

If you want a broad abstract analysis, Tom Slee had a good article on diversity in an Amazon-style recommendation system.  What I've noticed is a set of concrete examples of issues.

1. Amazon doesn't use negative valuations.  If it recommends 15 books by a given author and I mark 14 as "not interested", I still have to mark the 15th as "not interested" as well.  As  far as I can tell, low ratings for books don't degrade rankings of books tightly connected within the system, either. However, most books which are problems are not books one knows and loathes, but simply books which one has no interest in. (Generating negative evaluations out of only positive input doesn't always work: the LibraryThing Unsuggester works reasonably well on whole libraries, but I've found that it craps out badly on individual titles.)
2. It weights books based on orders which are explicitly gifts as it does normal orders.
3. Recommendations which seem to be most reliable in the sense that there's a really high correlation between a trigger and the recommendation seem to be the least useful ones.  Let me give a couple of examples to explain what I mean: I'm a professional software developer.  If I order, or indicate that I own, a technical book of core interest for me (say, for example,  Lakos' Large Scale C++ Software Design) the recommendations which are triggered have a strong likelihood to be of interest; unfortunately, they are also likely to be books I already have or know about.  The same thing is true in any well-defined academic area -- everybody in the area buys (or at least reads) the same core subset of books, but those people already know what that core subset is. In contrast, recommendations triggered by A Dance With Dragons are so broadly scattered over recent fantasy as to be useless, unless I were a singularly undiscriminating consumer of EFP. If I order a fiction book by an author I haven't ordered before, I usually have to prune a large number of recommendations I'm not interested in from the list shortly after.
4. And, look, If I indicate that I like a book by one author, I don't need help in finding other books by the same author.  There's no value-added there: all I need to do is a simple search. The vast majority of the books in my (pruned) recommendations list are by authors one or more of whose books I already have.
5. There's a special problem with children's books: they don't age the recommendations.  If I ordered a book for a six-year-old four years ago, the odds that I'm still interested in the same category is low: you should be recommending books for ten-year-olds.

As a practical matter, almost none of the books that I've bought over the last year have been ones to which I've been alerted by the recommendations.  Some I've bought as a result of online reviews or offhand favourable mentions by reviewers whose taste I trust in non-Amazon fora (online or print, notably the TLS); some as a result of browsing and seeing what has close physical proximity in a bricks-and-mortar bookstore (it's much easier to "look inside", especially at a truly random page or set of pages, when the book's in front of one, as well); some as a result of face-to-face recommendations.

The one area I've sometimes found Amazon useful is in popular science books in disciplines in which I'm not a specialist but am mildly interested. I think Melvin Konner's The Evolution of Childhood originally showed up as an Amazon recommendation.  But the useful recommendations are few and far between.