摘要： 分析NPA训练算法，指出其不足并提出改进措施。在第1类子循环的前半阶段采用Gilbert迭代，后半阶段采用NPA迭代，并提出界定这2个阶段的方法，利用中间计算结果优化了第2类子循环中的迭代过程。在不增加计算量的条件下，提高了算法收敛速度。基于该算法开发的自动分类模拟系统获得了较好的分类结果。

关键词: 数据挖掘, 支持向量机, NPA算法, 分类

Abstract: This paper investigates the NPA algorithm suggested by Keerthi et al commonly used in support vector training, points out its defects, and presents a modification. It takes Gilbert iteration in the early stage of type-Ⅰloop and NPA iteration in the later stage, and proposes a strategy of how to determine these two stages. It optimizes the NPA iteration in type-Ⅱloop by employing previously obtained computation results. In comparison with NPA, the modified version results in a remarkably accelerate convergence without increasing in computational complexity. By virtue of it, an experimental automated classifier is developed, which shows satisfactory performance.

Key words: data mining, support vector machine, NPA algorithm, classification

中图分类号: 

• TP181