Sp otlight ID #: W1

One-Pass Bo osting
Z. Barutcuoglu (Princeton), P.M. Long (Go ogle), R.A. Servedio (Columbia)

Motivation. Common b o osting scenario: at each stage the b est-p erforming base classifier is chosen from a fixed p o ol of candidates. If p o ol is very large (n-grams, amino acid sequences), may b e to o exp ensive to optimize at each stage. This pap er: b o osting algorithms that make one pass over p o ol of base classifiers · In one-pass setting it can pay to b e picky · For "diverse base classifiers" scenario, we get same p erformance guarantee as (more computationally intensive) multipass b o osting using one-pass algorithm · Exp erimental results