软间隔组合凸线性感知器设计

摘要

组合凸线性感知器是用来构造分片线性分类器的一个通用理论框架。对于凸可分和叠可分情况，分别使用支持凸线性感知器算法和支持组合凸线性感知器算法将两类样本分开。在此基础上，文中提出一种软间隔的组合凸线性感知器设计方法。该方法首先映射原空间数据到高维特征空间，然后利用K均值算法将其中一类样本聚类成多个簇，并在每一簇与另一类样本间构造凸线性感知器，最后集成组合凸线性感知器。该方法能解决原感知器模型不适用非叠可分数据的问题，并且在一定程度上简化模型结构，在保证分类精度的前提下，提高泛化能力。实验结果证实文中方法的有效性，同其它分片线性分类器的对比也说明了它的优势。%Multiconlitron is a general framework for constructing piecewise linear classifiers. For the convexly separable and the commonly separable datasets, it can correctly separate them by using support conlitron algorithm( SCA) and multiconlition algorithm( SMA) , respectively. On this basis, a soft margin method for multiconlitron design is proposed. Firstly, the training samples are mapped from the input space to a high dimensional feature space, and one class of those samples is clustered into some groups by K-means algorithm. Then, the conlitron is constructed between each group and another class of samples, and the integrated model, multiconlitron, is obtained. The proposed method can overcome the inapplicability of the original model to commonly inseparable datasets. By simplifying the model structure, the proposed method further improves the classification accuracy and the generalization ability. Experimental results show that the proposed method achieves better performance compared with some other piecewise linear classifiers and its effectiveness and advantages are verified.