A Study On Techniques Of Hierarchy Classification And Incremental Learning

In order to improve the performance of SVM training and testing of massive text samples in large categories, the effective solutions of hierarchy classification and incremental learning are researched in this paper.This paper first studies the hierarchy classification applied to the testing of large categories, and presents the solutions to the problems of constructing training texts, key feature lists and classifiers. Moreover, an incremental learning method, HISVML, is designed on the basis of hierarchy classification for the training of massive samples. HISVML includes the incremental updating of classification tree and feature set. While the incremental updating of classification tree is an on-line learning process according to the KKT path condition of new samples, and the incremental updating of feature set is a batch feature learning process before the incremental updating of classification tree. Finally, a SVM categorization system, RCCata, is proposed associating the hierarchy classification and HISVML methods in C++.The experiments on RCCata demonstrate that the hierarchy classification is superior in precision and efficiency comparing with the traditional single-level classification. Besides, less memory is required by RCCata during incremental learning in comparison with one-time learning.