多代表点的子空间分类算法

摘要

The multi-representatives nearest neighbor classifier, which builds classification model using model clusters centered with representatives, and determines the number of nearest neighbors automatically, has been proposed to overcome the shortcomings of traditional nearest neighbor algorithms. However, it would increase the number of model clusters when the samples in different categories are overlapped, and subsequently the prediction accuracy is affected. This paper proposes a multi-representatives-based algorithm for subspace classification, where the training samples are projected onto some different subspaces in order to construct the classification model consisting of model clusters in individual subspaces. This method makes the overlapped samples belonging to different classes in the entire space easily separable, so that the classification performances can be improved. In comparison with other methods such as traditional kNN (k nearest neighbor), kNNModel, SVM (support vector machine), etc., the experimental results show that the proposed method significantly improves the accuracy of the classification on datasets with complex category structures.%多代表点近邻分类克服了传统近邻分类算法的缺点,使用以代表点为中心的模型簇构造分类模型并自动确定近邻数目.此类算法在不同类别的样本存在大量重叠时将导致模型簇数量增大,造成预测精度下降.提出了一种多代表点的子空间分类算法,将不同类别的训练样本投影到多个不同的子空间,使用子空间模型簇构造分类模型,有效分隔了不同类别样本在全空间中重叠的区域,以提高分类性能.与传统的kNN(k nearest neighbor)、kNNModel、SVM(support vector machine)等分类算法的实验对比结果表明,新方法可以对复杂类别结构数据进行有效分类,且较好地提高了分类精度.