Machine Learning Meetup Notes: 2010-06-09

From Noisebridge
Jump to navigation Jump to search

Andy's final features:

  • problem hierarchy - 200 nominal
  • iq - percentage of getting it right on the first try, numeric
  • iq strength - number of those, numeric
  • chance - numeric
  • chance strength - how hard the step is - percentage of students who got the step right, numeric
  • num subskills - numeric
  • num tracedskills - numeric
  • predict -> cfa - percentage

all the numerics were discretized and run through naive bayes, that gave a 0.38 root mean squre error (MSE)

he stuck in a .87 for all of them and got a 0.33 MSE

final method: cfa = iq + 2(chance)/3 gave 0.31 MSE


Mike: presents clustering with graphs

  • JUNG - java tool for graph viz and contruction
  • eliminate edges that are the heaviest
  • minimum spanning tree
  • nb-sourceforge project
  • weight represents euclidean distance of two points
  • similar to pricipal component analysis