Two Stages Based Adaptive Sampling Boosting Method

To improve learning performance of boosting method, boosting learning procedure is divided into two sequential stages which are named reducing fitting error stage and reducing variance stage according to the idea of generalization error of boosting method composed of bias and variance which was originally proposed by Breiman. Traditional sampling methods such as roulette wheel selection is suitable for the learning procedure of reducing fitting error stage, and based on the characteristics of reducing variance stage, a new sampling method is proposed named SS method. Based on CSP and SS sampling methods, a new two stages based adaptive sampling boosting method named ASSBoosting is proposed, which according to the different characteristics of the two learning stages adaptively adopts sampling methods by the comparison of the prediction error caused by the two methods. The results of simulation confirm these findings