The Study Of Classification Based On Incremental Learning

This paper studies the incremental learning algorithm of support vector machine(SVM).The Statistical Learning Theory (SLT) is a new technique for solving various machine learning problems and shows that it is suitable for the finite data. Support Vector Machine (SVM) is a new machine-learning technique which was put forward by Vapnik in the last century. Support Vector Machine is based on statistical learning theory and provides a new theory for the finite data.SVM has some advantages including simple structure and good generalization, and been widely applied to many areas, such as pattern recognition, signal processing, automation, and data mining. However, the traditional SVM algorithm does not support incremental learning, which leads to its bad performance in large-scale data situation. So it has great significance in improving the classification performance of the incremental SVM algorithm both in theory and application fields.Now some famous SVM training algorithms have gotten extensive applications, such as chunking algorithm, decomposing algorithm, and SMO algorithm, etc. The main target of this paper is to find new incremental learning algorithms based on the current SVM algorithms, and this new algorithm will have a better classification performance in large-scale data situation, and the analysis of the algorithm is tested by experiments.Firstly, this paper introduces the mathematical basis of the least square support vector machine, analyzes the characteristics of support vectors and the processing of incremental learning, then presents a new algorithm of incremental learning. This algorithm uses the prediction error threshold to retain the useful information to decrease sample training scale. The experiment based on UCI data sets proves the algorithm can obtain a faster training rate and higher classification accuracy.Secondly, this paper oroposes a new algorithm of incremental learning based on membership function. In the process of incremental learning,the wrong support vector can be transformed to support vector. The algorithm uses a membership function of the wrong support vector to decrease sample training scale. The experiment based on UCI data sets proves the algorithm is feasible.