一种基于聚类核的半监督支持向量机分类方法

为了在标记样本数目有限时尽可能地提高支持向量机的分类精度,提出了一种基于聚类核的半监督支持向量机分类方法。该算法依据聚类假设,即属于同一类的样本点在聚类中被分为同一类的可能性较大的原则去对核函数进行构造。采用K-均值聚类算法对已有的标记样本和所有的无标记样本进行多次聚类,根据最终的聚类结果去构造聚类核函数,从而更好地反映样本间的相似程度,然后将其用于支持向量机的训练和分类。理论分析和计算机仿真结果表明,该方法充分利用了无标记样本信息,提高了支持向量机的分类精度。

Semi-supervised SVM classification method based on cluster kernel

In order to improve the classification accuracy of support vector machine when limited the labeled samples, this paper proposed a semi-supervised support vector machine classification method. It constructed the kernel function according to the cluster assumption. The method used both the labeled and the unlabeled samples to construct the kernel function by the K-means algorithm. The similarity between the samples could be represented better by such kernel function. Then it used in SVM to train and obtain the classification results. Theoretical analysis and computer simulation results show that the algorithm can effectively use a large number of unlabeled samples, and can improve the classification accuracy.