A Two-Level Incremental Training Date Algorithm For Auto SVM Model

Noisy or uncertain data are common in machine learning and data mining applications. Noisy data can significantly affect the robustness and effectiveness of the data mining and machine learning algorithms. To eliminate the negative affects, A new classify algorithm is proposed for SVM model based on the standard SVM algorithm.Generally speaking, the penalty parameter of the standard SVM model is hold to a constant during the whole classify process. The fact that the penalty parameter does not vary according to the discovery of the data properties is inappropriate. So, on the basis of the previous works, a dual quadratic programming model is proposed in which the penalty parameter varied. Then, the existence of the optimal solutions of the Gauss kernel SVM model is proved. According to the existence, a new algorithm for the SVM model is proposed, which constructed on the incremental of the training data set. The process of the algorithm and some cautions when select the penalty parameter is given.The results of the numerical experiments prove that the new algorithm has the robustness to the number of the data properties. When the times of the iterations increase, the accuracy of the algorithm has better convergence. This means that the new algorithm can handle the noisy data better.