Adaptive Boosting

back to notes

Boosting

Boosting

AdaBoost (Freund, Schapire)

Adaboost stands for adaptive boosting.

Given:

   classified samples: (x_0,0,x_0,1,...x_0,n, c₀), (x_1,0,x_1,1,...x_1,n, c₁), ..., (x_N,0,x_N,1,...x_N,n, c_N)

   where the classes c_i are +1 or  -1    (i=1..N)

   initial weights are starting all equal:  w_0,i = 1/N    (i=1..N)

Algorithm:  for training iterations t (t=1..T) for all i classifiers (i=1..N):

   1) Get all the classes back from the classifiers: h_i     

   2) Calculate the error:  E = P(h_i!=c_i)

   3) Calculate alpha:

   4) Update the weights:

   5) Result: 

• Another short intro re. AdaBoost: MlNet/OIS explanation