In AdaBoost, a rescaling factor multiplies the weights of misclassified points so later weak learners give them more importance. It is the total weight of correctly classified points divided by the total weight of incorrectly classified points. section 12.3
Two rounds of reweighting
With ten equally weighted points, a weak learner classifies seven correctly and three incorrectly, giving a factor of 7/3 ≈ 2.33. In the next round, the correct points’ weights total 11 and the incorrect points’ weights total 3, so the factor is 11/3 ≈ 3.67. section 12.3
Figure 12.10 Fitting the first weak learner of the AdaBoost model. Left: The dataset, where each point gets assigned a weight of 1. Middle: A weak learner that best fits this dataset. Right: The rescaled dataset, where we have enlarged the misclassified points by a rescaling factor of 7/3.
Compare the original dataset, the weak learner’s split, and the enlarged weights on misclassified points. section 12.3
