β Input: |
β : training set consisting of instances; |
β : base learner whose output is assumed to be a class probability distribution; |
β : desired ensemble size; |
β : maximum number of iterations to construct an ensemble classifier; |
β : a factor to determine number of artificial instances to generate. |
β Training phase |
ββββ Initialization: |
ββββLet and ; |
ββββProvide the given training set as the input of base learner to get a classifier ; |
ββββInitialize ensemble set ; |
ββββββ Compute ensemble error as |
ββββββββββββββ. ββββ(3) |
βββ While and |
ββ β(1) Generate training instances, , according to the distribution of training data; |
β β β(2) Label each instance in with probabilities that each class label is selected |
βββββbeing inversely proportional to those predicted by ; |
β β β(3) Combine with to get a new training set ; |
β β β(4) Apply base learner to to obtain a new classifier ; |
β β ββ(5) Add to ensemble set , namely, let ; |
β β ββ(6) Based on the training set , compute the ensemble error of , say,, as |
ββββ that done in equation (3); |
β β ββ(7) If , let and update ensemble error as ; Otherwise, |
ββββdelete from the ensemble set , that is, ; |
ββ ββ(8) ; |
βββ EndWhile |
β Prediction phase |
βββ Let be the probability that comes from class supported by the classifier . |
ββCalculate the confidence for each class by the mean combination rule, that is, |
βββββββββββ,βββ(4) |
ββwhere stands for the real ensemble size. |
βββ Assign to the class with the largest confidence. |