Research On Dynamic Construction Of Incremental Support Vector Machine Based On Clustering

The data stream is produced continuously with high speed, which makes the traditional data mining approach ineffective because the original data sets are iteratively scanned. How to apply support vector machine which is characterized with simple structure, global optimization and good generalization in data stream mining has become a hot research, and the complexity of support vector machine becomes the "bottleneck" problem in dealing with the large-scale data set.In order to deal with the data stream mining with incremental support vector machine, a dynamic construction algorithm of incremental support vector machine based on clustering is proposed after several incremental support vector machine algorithms and the characteristics of the data stream based on the statistic learning theory and support vector machine is analyzed. Firstly, the K-means clustering method is used to regulate the training data set of the incremental learning, aiming at reducing the sample distribution difference of the same data set, and improving the sample distribution difference of the different data sets. This is able to aggravate the characteristic of the dynamic data stream and improve the performance of algorithm. Secondly, after the reason that the existed classifiers combination algorithms are not suitable for multiple support vector machines combination is analyzed, the multiple Support vector machines combination algorithm based on clustering partition is proposed. The samples that are correctly classified by some classifiers but mistakenly classified by other classifiers are picked out to deal with solely. The feature space is divided according to the clustering results, which is similar to the construction process of C-MCC and the classifier which has the optimal performance is selected as the final output of system.Simulation results show that the dynamic construction algorithm of incremental support vector machine has a better classification performance compared with the traditional w-model incremental learning algorithm. And at the same time, the proposed multiple support vector machines combination algorithm based on clustering partition has better combination precision than Voting, K-NN and C-MCC.