correlate Rotating Header Image

Negatives of Recommendation Systems

Jun 19th, 2008
by Lou Paglia.

For the most part, I’ve been impressed with Amazon’s e-commerce capabilities: reviews, book recommendations and very often, their algorithmic book bundling system. However, I have now witnessed one that has gone terribly wrong. My wife bought the book on tape, The Lady and The Panda.

In this particular case, Amazon’s algorithm went terribly ary:

Check out the bottom of the graphic above, and the book that is recommended: Goddesses, Whores, Wives and Slaves.  Their scoring mechanism missed by quite a big margin on this one.

Sphere: Related Content

Posted in: Uncorrelated.
Tagged: · ·

← Metrics-Driven Innovation
Google Heads →

Viewing 2 Comments

    • ^
    • v
    At first blush, I don't see the problem.
    What would you say is the nature of the problem?
    And how might the matching algorithm be improved?

    Why should a matching algorithm consider the book "Goddesses, Whores, Wives and Slaves" a poor complement to "The Lady and the Panda"?

    The Editorial Reviews for both books allude to feminist readings of contemporary history in the case of the "Panda" and antiquity in the case of "Goddesses".

    For example, a snippet from the Editorial Review of the "Panda":
    "The exploits of the utterly macho men who bagged the beasts also made good adventure-film fodder. Yet one of the most famous animals ever brought to America—the giant panda—was captured by a woman, Ruth Harkness."

    And here's a snippet from the Editorial Review of "Goddesses":
    "The first general treatment of women in the ancient world to reflect the critical insights of modern feminism."

    Looking at those reviews I can see how the algorithm may have functioned to select these items as complementary.

    So the above points to why the book was selected as a complement. The question now is, why should it be excluded after initial selection?

    Is it the title? Provocative no doubt in its allusion to inflammatory archetypes. That would be tricky to do though: identify mores of your customer based on past titles selected for purchase.

    Or is it that one book focuses on Contemporary History and the other Antiquity?

    You know in the end perhaps the recommendation feature works best if it is coupled with a binary input for the customer. "Yes that is a good recommendation", "No that isn't".

    If it is an issue of mores; that subjective scale is best attacked via direct user input.

    Right now the only way to determine if the customer concurs with the recommendation is if she purchases the recommended complement.
    • ^
    • v
    Jeremiah: You bring up good points as always. Here's my additional take: First off, I "pick" on Amazon because they have one of the best bundling strategies and recommendation systems around. They've been perfecting it from the late 90's. And let's be honest, I believe I may have been wrong with the word "recommendation", their bundling is most likely tied to related books and which ones in inventory they have a deal or are trying to unload.

    I think the fundamental issue is recommendation/bundling relevancy to the customer. Your view of the the context of the books is actually very similar is an important one, but not to my wife. My wife was reading the Panda book with a different goal than one that she is interested in feminist literature. So that isn't the only way to look at recommendation or bundling.

    Often times the books I buy as bundles is purely a fiscal decision. I can get two "recommended" books together at the same time for a better price.

    And on another topic, the recommendation is in the eye of the beholder. Perhaps 99 of 100 people who saw that page would have immediately since the contextual relationship of the two books. I simply may have been the one that did not and thought it merited the conversation so thanks for engaging it!

    See you in the office. :)
 

Trackbacks

(Trackback URL)

close Reblog this comment
Powered by Disqus · Learn more
blog comments powered by Disqus