In adaboost among the weight of the sample weight alpha is fixed, where the five-pointed star ring blue ○ in 3 minutes wrong, where the five-pointed star in a red circle 4 × 1 ○ and have a pair of points, it is easy to I think that this model, red for the position of judge more credible.
Dynamic weight, each x will have a specific weight, different classifiers for different weights of the weight of the sample is not the same
classifer Base : min ○ and △
Predictor Competency : points and points to the wrong sample
The misclassification normalized to a method of use of KNN, for example, to test a sample xi and five nearest training samples to calculate the model or wrong in the points five samples above right, if that's five points are on the that this model more reliable.
But Euclidean distance in high-dimensional space
L1: Manhattan distance
L0.5: fractional distance
As shown above, Regionboost poor convergence, adaboost good convergence
regionboost measurement errors to below adaboost
Corresponding Author:
references: