Another Amazon recommendation ... this one associates Snakes on a Plane and World Trade Center.  The only thing these movies have in common is the release date.

 

Comments:

Well frankly, if you buy either of those, you deserve to have to sit through the other...

Posted by Calum on May 28, 2007 at 08:49 PM EDT #

Hi, Paul. Picking on Amazon a lot lately, aren't you? :) On the broader issue here, you appear to be suggesting that Amazon would have much better recommendations if they only used some kind of content filtering or content-based clustering to restrict to items in the same general category. I am not convinced that is correct. The problem with restricting based on content is that it makes the recommendations obvious and restricts serendipity. For example, recommending Batman movies for a Spiderman DVD or other Samuel Jackson movies for Snakes on a Plane DVD is not surprising or all that interesting. Those are obvious recommendations of things you easily could find on your own. The value of recommendations is helping people discover things they might not have found on their own. That requires reaching beyond the obvious and into the surprising. More mistakes might be made by doing that -- and you appear to have highlighted a selection of mistakes in your recent posts -- but the overall impact often is more useful and more effective recommendations.

Posted by Greg Linden on May 29, 2007 at 11:13 AM EDT #

Greg - Don't get me wrong, I like Amazon's recommender quite a bit .. the last time I shopped at Amazon (last week), I went to buy one book, and ended up buying three, on the strength of the recommendations and supported by customer reviews. I think Amazon's rec engine is probably the best out there ... nevertheless, there are times when it falls short, and as typical with CF systems, when the recommender goes awry it yields puzzling or humorous results such as the Narnia and Nose hair example (I still find myself lying awake at night mumbling to myself "Narnia, nose hair .. what is the connection?").

Indeed, serendipity and novelness are traits of a good recommender - most of the music recommenders I've seen lack that trait completely, I don't need a recommendation for The beatles or Radiohead, thank you very much.

I don't necessarily think that content-based filtering is called for to restrict items ... indeed I like the idea of cross genre recommendations - like 'the matrix', you may like this trench coat.

These posts are a way for me to keep track of the common types of mistakes that we see in recommenders. I use the Amazon recommendations as examples because everyone is familiar with these type of recommendations and there are plenty of examples to draw from. The goal is to get a better understanding of what people are looking for in a recommender. Novelness and serendipity are clearly important, but so are transparency and trust. If I get a recommendation, I like to know why something is being recommended ... and I like to know that the recommendation is not biased by the mob or by money.

Posted by Paul on May 29, 2007 at 01:07 PM EDT #

Post a Comment:
Comments are closed for this entry.

This blog copyright 2010 by plamere