利用相似度分割特征集的混合核构造方法

Mixed Kernel Construction Method Based on Feature Set Division

核函数作为样本相似性的衡量尺度是影响支持向量机分类效果的重要因素。为了使相似性衡量尺度与样本特征的分布特点相适应,提出利用相似度分割特征集的混合核函数构造方法。通过研究每维特征在不同相似性函数下的区分能力,将特征集分割成相应的特征子集,并根据特征子集在分类中的重要性程度,对相应的样本相似度矩阵进行线性加权,得到最终的分类结果。实验结果表明,文中提出的混合核可以同时具有多种不同特性核函数的优点,改善了支持向量机分类器性能。

As the similarity metrics of the samples, kernel function is a key factor that affects the classification. In order to adjust the similarity metrics to the distribution of the feature, a kernel function construction based on feature set division is proposed. The feature set can be divided into several different subsets corresponding to the distinguish ability of the feature and then by linear weighing the similarity matrixes according to their importance, the classification can be realized. The algorithm proposed possesses the virtues of different kernel function and has a better performance.