近邻分类法的改进及其在刀具分类中的应用

SOME IMPROVMENTS OF K-NN CLASSIFICATION AND ITS APPLICATION ON THE LATHE TOOLS

本文研究三种在近邻分类法基础上的改进算法:(1)编辑技术,(2)边界抽取,(3)边界补缀。编辑技术使识别率已接近Bayes 标准。再经边界抽取仅保存对分类直接起作用的设计样本集,从而提高了分类速度,但识别率有所下降。边界补缀法是本文作者之一最近提出的一种新算法。此算法在重新审查部分设计样本的过程中,“拣”回在边界抽取时失掉的个别位于两类交界处的样本,进一步完善了分类器。补缀后的边界样本集使分类速度保持在(2)的水平,而识别率又上升到使用编辑技术(1)后的情况。将以上算法依次应用到刀具分类的两类问题,给出可喜的结果。经过(1)、(2)、(3)算法处理过的设计样本集,其识别率接近大样本集时1-NNR 的结果,同时分类速度也提高了约6倍。

This paper analysis three different kinds of improved algorithms based onthe K-NN classification:(1) Editing technique,(2) Boundary extraction,(3) Boundarypatching.The editing technique makes the error rate very close to Bayes standard.After the boundary extraction,only the boundary samples of each class is saved indesign set as a new one,thus the classification speeds up,out the error rate increases aswell.The third improvment boundary patching is a new algorithm proposed by one ofthe authors of this paper.It examines a part of the design set formed in editing andpicks up some samples lost during boundary extraction,it futher improves theclassifier.The speed of the classifier is still in the lovel of (2),but the error rate hasbeen dropped down to Bayes standard.Applying the algorithms above to the two classes of the problems of lathe toolclassification,the result is quite effective.By using the algorithm (1),(2),(3),thespeed of 1-NNR classification has raised 6 times as much fast as before and almost nodifference with Bayes standard on the error rate.