Monday, October 8, 2012

how many neighbors? normalize?


  • N=1 : normalize nums min..max 0..1
  • k=K : use mean of kth nearest neighbors
  • data sets : desharnais, autompg, nasa93, finnish, miyazaki94, china, coc81
  • results : mre
  • small effect size : median of numbers bottom third (max(mre) - min(mre))/3


results: http://unbox.org/stuff/var/timm/12/dash/var/egknn.out

note the 010101 patterns on everything except china which is 1111110000000

  • normally normalization does not matter
  • but when it matters, it really matters


in the n=1 results

  • k=1 has less error in 6/7 data sets (exception: desharnais)

conclusion: k=1 a good default

No comments:

Post a Comment