3 comments

[ 3.0 ms ] story [ 15.0 ms ] thread
Was just about to say that it doesn't work well with large lists of words because at some point the gaps between words are small enough there's a path right through the entire set which causes them to be "clustered".

  holden
  golden
  golder
  colder
  corder
  border
  bonder
  wonder
  wander
  warder
But I see that once you get longer strings then it works great. I like the log message processing.
Edit distance for syslog message classification. Interesting!

One notable caveat is that there's no clear axis of classification. Sometimes things may be classified together because they talk about the same resource (with a long name). Sometimes things may be classified together because they have the same message template. Sometimes thing will be classified together because they share a lot of digits of time (not necessarily close in time).

Odds are you get some of each type of clustering.