Research of Incremental Learning Algorithm for SVM Based on Class Center Diameter

Abstract

Abstract In the learning process based on ISVM, how to effectively retain the history information and selectively discard some new training data, so as to maintain the classification accuracy and save the storage space after adding new samples every time, which is the key of the current ISVM classification algorithm. This thesis proposes a new incremental learning algorithm, that is, SVM incremental learning algorithm based on cluster diameter (CD-ISVM). This algorithm firstly calculates the two centers of the positive and negative samples, and then to build the boundary vector set by the coordinates of the two class centers. Moreover, KKT conditions are combined to filter incremental data and boundary vector set, and feedback information is provided to adjust the position of the boundary vector set. Then, the union of SV set and boundary vector set is taken as the increment set. After several increments, a classifier with strong generalization is finally trained.