Tag: machine learning

June 10, 2004

  • Dumb Smart Search

    I was just trying out the new version of Napster and when I searched for “China Crisis” I got results that included David Sylvian and Asia albums.

    Why might that be?

    Well, David Sylvian used to be in a band called “Japan” which is close to China and “Asia” is where China is.

    I’m surprised I didn’t get “Dishwalla” as one of the results!

    Napster seems to have some sort of “concept” search algorithms in use that really doesn’t make sense in this context. Given that Napster knows I’m searching for an artist, it seems that there are two approaches to expanding search beyond the original term, “spelling” and “related projects”.

    There are band and artist names that are hard to spell or that you only vaguely remember from youth (was “Hitchin’ A Ride” done by Vanity Fair or Vanity Fare?). In this case implementing something like Google’s “Did you mean…?” feature would be very smart. I want this kind of help so I don’t have to remember how to spell Alanis Morissette (Napster catches typos on her last name).

    If I’m searching for “Tin Machine” it might be useful to offer results for frontman David Bowie as well. “Related Projects” searches could be very helpful particularly when you remember David Byrne singing some song but you don’t realize that it was from a solo album not a Talking Head disc.

    The problem with Napster is a search on Tin Machine produces “Tony McKinney”, “More Machine Than Man”, “Nick Gilder and Time Machine” and (very oddly) “The Clancy Brothers & Tommy Makem”. These are guesses at possible typos but because they aren’t identified as such it looks like they have a really bad search feature.

    Lessons Learned:

    1. Smart search is dumb if it does not take into account the user’s goal in doing the search in the first place.

    2. Tell the user why you are presenting results that are not expected (“No matches found for Tin Machine. Did you mean…? Artists related to Tin Machine include…”).

    3. Hard code results for very popular searches so you can give really relevant information.

October 17, 2003

  • Thought

    Jupiter Research: Web Site “Personalization” Does Not Always Provide Positive Results:

    “According to Matthew Berk, Research Director at Jupiter Research, ‘Most Web site personalization projects fail to deliver real business benefits. Our industry has always assumed that a personalized Web site was a better one, both for the visitor and the site operator. Our research has found that this is not the case.’

    According to the report, for every intended benefit tied to a personalization-related agenda, site operators can select from many other tactics to achieve the same goals, at far lower cost. “To drive key business metrics, most sites are better off focusing on the basics, like usability, information architecture and making key tasks easy for users to accomplish,” said David Schatsky, Senior Vice President at Jupiter Research.”

August 8, 2003

  • Thought

    The problem with contextual advertising is that computers don’t really understand context, they just look like they do.

    The most common way this shows up is in ads that match keywords, but not the intent of the page.

    For example, amazon.com’s page for

    Alfie Cohen’s “Punished By Rewards: The Trouble with Gold Stars, Incentive Plans, A’s, Praise, and Other Bribes” offers these contextual ads (served by Google):

    Incentive Toolkit

    The Tools You Need For Your Next Incentive

    www.incentivetoolkit.com

    Incentive Plans

    Find Solutions for Your Business Free Reports, Info. & Registration

    www.KnowledgeStorm.com

    Award Programs

    Personalized awards & incentives sold with our Price Guarantee.

    www.custom-promotional-logo.com

    Oh dear. These folks won’t get a very good response selling ads for incentive programs on a page for a book that is explicitly AGAINST incentives.

August 17, 2002

  • Thought

    A good New York Times article on the foibles of voice recognition.

    For example, here is a list of “wordos” that author David Pogue’s software created (some quite funny):

    bookmark it -> book market
    Motorola -> motor roll a
    modem port -> mode import
    a procedure -> upper seizure
    and then stick it in the mail -> and dense thicket in the mail
    movie clips -> move eclipse
    I might add -> I my dad
    inscrutable -> in screw double
    hyphenate -> -8
    suffocate -> Suffolk 8
    a case we summarily dismissed -> a case we so merrily dismissed
    or take a shower -> Ortega shower
    the right or left -> the writer left
    oxymoron -> ax a moron
    ArialPhone guy -> aerial fungi

    Still, I can’t help thinking that voice input is inevitable — as are translation errors. I wonder if it is possible that humans will modify pronunciations to accommodate the machines. This would be analogous to Newton failing in writing recognition but the Palm succeeding because it used “Graffiti” a made up alphabet. It was easier for the humans to learn to deal with the ambiguity than the machines.