Boosting: Margins

• What is boosting trying to do?
  • Maximize the margin for each training example
  • For each example ex
    • Initialize sumex to zero
    • For each hypothesis H in the hypothesis list
      • If H correctly classifies ex, add its weight to sumex
      • If H incorrectly classifies ex, subtract its weight from sumex
  • The margin for ex is sumex / sum(all weights of H)
• What is the range of the margin?
• What is the meaning of the margin?