The MIR propellerheads of the world are gearing up for ISMIR 2007 later this week, the uber conference for music similarity geeks.  Seeing as the timing is right, I thought I would present an example of why this research can be difficult.  I’m interested in hearing any thoughts on the example - the good, the bad and the ugly, and if you have a better one.

The purpose of a great deal of similarity work is to generate music recommendations.  This only complicates the matter further as similarity starts to mean “what people will like”.  Even if we ignore this precarious assumption and its untold consequences, there are still many different reasons why people may believe that music is similar.  This makes it very difficult to measure the effectiveness of any similarity algorithm.  For example, check out this playlist:

http://new.music.yahoo.com/playlist/1CEC1111-F9D6-409D-BCB6-53558D25506C

It contains three different versions of the song All Along the Watchtower - one by Bob Dylan, the other by Jimi Hendrix and finally a more recent cover by U2.  Which two are more similar?

U2’s version is essentially an update of the Jimi Hendrix rendition and features very similar instrumentation.  The original Dylan tune is stripped down and folky.  It sounds different from either the Hendrix remake or the U2 cover.  Yet Dylan and Jimi recorded their tunes 20 years before U2 and the two (Hendrix and Dylan) will be forever linked together as icons of classic rock and the 60s.

I think it’s a nice example because it shows the difference in people’s perception of what similarity can mean from one context to the next.  Is it that they sound alike?  Is it that you think of them as a belonging to a common group?  Is it that they are using similar chords or instrumentation?  Or is it that you fondly remember listening to them together when you were 16?

In the end, music similarity seems to be different things to different people at different times.

Popularity: 15% [?]