一种新的动态SVM选择集成算法

针对动态选择集成算法存在当局部分类器无法对待测样本正确分类时避免错分的问题,提出基于差异聚类的动态SVM选择集成算法。算法首先对训练样本实施聚类,对于每个聚类,算法根据精度及差异度选择合适的分类器进行集成,并根据这些分类器集成结果为每个聚类标定错分样本区,同时额外为之设计一组分类器集合。在测试过程中,根据待测样本所属子聚类及在子聚类中离错分样本区的远近,选择合适的分类器集合为之分类,尽最大可能的减少由上一问题所带来的盲区。在UCI数据集上与Bagging-SVM算法及文献［10］所提算法比较,使用该算法在保证测试速度的同时,能有效提高分类精度。

A New Dynamic SVM Selected Ensemble Algorithm

Dynamic Selection of integration algorithm is usually accompanied with the situation that there is no way to avoid the misclassification when the local classifier can not classify the test pattern correctly, accordingly a novel dynamic SVM selection ensemble algorithm based on diversity-clustering is proposed. Clustering is applied to training samples firstly in this method. To every clustering, appropriate classifier ensemble is selected based on accuracy and diversity, and the sample areas which are misclassified by the classifier ensemble for every clustering is demarcated, and a set of classifier ensemble for it is designed. During testing, the test sample is classified by the appropriate classifier ensemble based on the clustering to which it belongs and the distance between it and the misclassified sample areas. Using this method can remarkably reduce the blind regions while the test sample is very close to the misclassified areas mentioned above. Experimental results show the effectiveness of this method. Compared with Bagging-SVM and literature \10\] on UCI data set, the testing speed can be guaranteed and simultaneously the classification accuracy can be effectively improved by using this algorithm.