6 Responses

  1. guillermo
    guillermo March 3, 2009 at 05:36 |

    How about populating music, movies, poetry, books, in general everything that we might like and has a sentimental value (which is just to say the things that imply preference based on interpretation) with a semantically rich recommendation system? Simply put, if you wanted a recommendation on a particular thing you should start by describing it.

    Take del.icio.us for example. In order for the service to be of any value I first have to tag the different websites I want to bookmark, and then I can see websites that match those bookmarks.

    Or better yet, Wikipedia, it’s the best music recommendation system I’ve used. Not because every artist I search for is connected to other artists I like, but simply because it directs me towards artists I _might_ like.

    It’s an interesting abstraction, the notion of a recommendation not as a finality in on itself, that is, not implying a choice, but rather as a tool that helps us make that choice.

  2. Thomas Kjelsrud
    Thomas Kjelsrud March 3, 2009 at 09:45 |

    Interesting read! I have no doubt that keeping the number of options down is beneficiary in several web site scenarios. However, I’m a bit troubled by this model within a larger commercial site as it would inherently limit competition.

  3. Vegard Sandvold
    Vegard Sandvold March 3, 2009 at 11:19 |

    @Thomas Kjelsrud
    I haven’t given the competition perspective much thought so far. I was more worried about how choice effects the consumer emotionally. But what you’re saying seems to be important.

    Exactly how do you think this bare-bone recommendation scheme will limit competition? Will it allow producers to raise the prices on their goods? Or do you think it may favorize some producers at the expense of others?

  4. Thomas Kjelsrud
    Thomas Kjelsrud March 3, 2009 at 13:34 |

    I think it will limit the chances of getting notices within a specific field or topic, as the well-established dominate these search & recommended results.

    One could of course ask if it is “noble” to e.g, publish yet another book on a popular topic in order to draw sales from “the hype”, or should they be “forced” to publish & write books on more novel areas in order to dominate their own clusters of knowledge (and keywords being searched).

    Although limiting any result to the top “N” can be beneficiary for users, the ones struggling to get noticed further down in the results should still have a chance to be seen.

    I guess to wrap it up: how do you monetize on the long tail while limiting the choices to the most popular/most recommended? (which is not really part of the long tail at all)

  5. Phoebe
    Phoebe March 3, 2009 at 14:24 |

    Guillermo’s comment about populating music, movies etc. with a semantically rich recommendation system is very interesting. This is exactly what we do at Jinni (http://www.jinni.com), the first semantic discovery engine for movies and TV shows.

    One advantage of the semantic approach is that it’s better for the long tail than statistical approaches that rely on what’s already popular with other people.

  6. Gene Golovchinsky
    Gene Golovchinsky March 14, 2009 at 23:25 |

    It makes sense to me to limit the choices to make the act of choosing less of a cognitive headache. To address Thomas Kjelsrud’s point about the long tail, one possible answer is that the long tail doesn’t belong on the same body as the big head.

    If the person’s information need (rather than the query that loosely approximates it) is taken into account, it may be possible to present results in a more appropriate manner than a single ranked list. Various clustering approaches may be used, for example, to identify categories of documents that are popular, unusual, detailed, technical, etc. Searchers could then use these categories (or many other possible ones) to help them make sense of search results in a more task-oriented manner.

    Another (related) approach is to think about aspects, and to factor search results that way. Some of these approaches are more difficult computationally, but for many domains there may be useful metadata that can be leveraged, or reasonable heuristics may be applied.

    The disadvantage of considering search results this way is that one size doesn’t necessarily fit all users, and the search provider must do more work to make the results more useful. Of course if you _can_ do it well, you’ve got a competitive advantage.

Comments are closed.