Exploring the Set of Sparse , Optimal Classifiers
Exploring the Set of Sparse , Optimal Classifiers
复制标题
探索稀疏最优分类器集
DOI：
--
复制
发表时间：
2003
期刊：
影响因子：
--
通讯作者：
Martin Brown
中科院分区：
文献类型：
--
作者：
Martin Brown
特征选择是分类器设计的重要组成部分。然而，确定经验分类器的适当结构是一个困难的过程，通常需要广泛的探索和评估。本文描述了一种新的方法，表示的特征选择过程作为一个正则化/优化问题，它有一个单一的全局最小值。此外，还描述了一种新算法，该算法允许设计者有效地构建完整的基于稀疏核的分类器家族，因此可以交互式探索它们的结构。这允许设计者在正则化参数被改变时研究参数的轨迹，并寻找诸如在许多多变量数据分析问题中发生的辛普森悖论的影响。该方法被证明是著名的澳大利亚信贷数据集。
Feature selection is an important part of classifier design. However, determining an appropriate structure for the empirical classifier is a difficult process that often requires extensive exploration and evaluation. This paper describes a novel approach that represents the feature selection process as a regularization/optimization problem, which has a single global minimum. In addition, a new algorithm is described that allows the designer to efficiently construct the complete family of sparse kernel-based classifiers, and therefore their structures can be explored interactively. This allows the designer to investigate the parameters’ trajectories as the regularization parameter is altered and look for effects such as Simpson’s paradox that occurs in many multivariate data analysis problems. The approach is demonstrated on the well-known Australian Credit data set.