高级检索
    田永军 陈松灿. 面向矩阵模式的正则化Ho-Kashyap算法[J]. 计算机研究与发展, 2005, 42(9): 1628-1632.
    引用本文: 田永军 陈松灿. 面向矩阵模式的正则化Ho-Kashyap算法[J]. 计算机研究与发展, 2005, 42(9): 1628-1632.
    Tian Yongjun and Chen Songcan. Matrix-Pattern-Oriented Ho-Kashyap Classifier with Regularization Learning[J]. Journal of Computer Research and Development, 2005, 42(9): 1628-1632.
    Citation: Tian Yongjun and Chen Songcan. Matrix-Pattern-Oriented Ho-Kashyap Classifier with Regularization Learning[J]. Journal of Computer Research and Development, 2005, 42(9): 1628-1632.

    面向矩阵模式的正则化Ho-Kashyap算法

    Matrix-Pattern-Oriented Ho-Kashyap Classifier with Regularization Learning

    • 摘要: 线性分类器由于其简单性和易扩展成非线性分类器的特性,使其成为统计模式识别中最常用的方法之一.正则化的Ho-Kashyap线性分类算法(MHKS)采用了支持向量机最大化间隔的思想.现有的线性分类器大都是针对向量模式的,要应用于矩阵表示的模式,如人脸图像等必须首先将矩阵模式转换成向量模式.但如此至少会带来3个不足:①原有矩阵模式的空间或结构信息可能会遭到破坏;②由于权向量的维数等于输入模式的维数,当输入模式维数很大时,权值的存储空间相应地会很大;③对于大维数的模式,当样本数不多时,利用线性分类器易导致过拟合.受到已有面向矩阵的特征提取方法的启发,设计出面向矩阵模式的双边正则化Ho-Kashyap分类算法MatMHKS,克服了以上不足.与MHKS相比,在ORL数据库、Letter数据集、UCI机器学习部分数据集上实验都取得了更好的分类性能.

       

      Abstract: Linear classifier is a common method in statistical pattern recognition. The modified Ho-kashyap with square approximation of the misclassification errors (MHKS) is a linear classifier designed similarly as the support vector machine to maximize the separation margin. However, the conventional linear classifiers are almost based on vector patterns, i.e., before applying them, any non-vector pattern should be first vectorized into a vector pattern. But, such a vectorization will bring three potential problems at least: ①Some implicit structural or contextual information may be corrupted; ②the higher the dimension of input pattern, the more space are needed for storing weight vector; ③When the dimension of pattern is very high and the sample size is small, linear classifier tends to be overstrained. In this paper, a new classifier design method based on matrix patterns, called two-sided linear classifier (MatMHKS), is proposed by modifying the MHKS algorithm. This method can overcome above problems. Experiment results on real dataset show that MatMHKS is more powerful than MHKS.

       

    /

    返回文章
    返回