交叉样例选择研究.doc

Classified Index: CODE: 10075
: NO: 20101192
A Dissertation for the Degree of M. Engineering
Research on Cross Sample Selection
Candidate: Li Ta
Superisor: Prof. Wang Xizhao
Prof. Zhai Junhai
Academic Degree Applied: Master of Engineering
Specialty: Computer Applied Technology
University: Hebei University
Date of Oral Examination: May, 2013
摘要
摘要
海量数据是数字时代明显的特征,数据大规模增长使得数据处理变得异常困难,样
例选择是处理海量数据的主要方法之一。样例选择的目的是去除原始数据集中的冗余样
例和噪音样例,从而得到一个能够代表原数据集的小规模子集。传统的样例选择算法大
多是针对某一种分类器而设计的,因此其结果往往也只适用于这种类型的分类器。此外,
随着数据集规模的增长,经典的样例选择算法难于处理大规模数据集,高效率的样例选
择算法,特别是针对大数据集的样例选择算法成为近几年的研究热点。
本文针对较大数据集分类问题,提出了能够用于不同分类器的样例选择算法,该算
法基于分治和交叉验证的思想进行样例选择。具体地,首先将数据集分为若干不相交的
子集,对于每一个数据集,用其他子集训练出的分类器组成委员会,对该子集进行样例
选择。针对输出为后验概率的分类器使用K‐L 散度度量样例的重要性,针对其他分类器
使用投票熵度量样例的重要性。当使用速度快的ELM 分类器时,本文提出的算法能够快
速处理较大规模的数据集。为了验证本文算法的有效性,在19 个数据集上进行了实验,
N、ENN、RNN、MCS 和ICF)在时间、精度和压缩比等方
面进行了实验比较。实验验证了交叉样例选择算法在多种分类器上应用的可行性,并能
够快速处理较大规模的数据集,得到压缩比更高的子集。
关键词样例选择投票熵 K‐L 散度 ELM 概率神经网络
I
Abstract
Abstract
Huge amounts of data is the outstanding characteristics of the digital age, and the data
processing es extremely difficult with the massive growth data, then the sample
selection is one of the main ways to deal with huge amounts of data. The main purpose of
sample selection is to remove the redundancy and noise samples from the original datasets, a
small representative subset of original dataset can be obtained. Most of the traditional sample
selection algorithms are tailored for a particular classifier, so the samples selected with these
algorithms are often only applicable to this type of classifiers. In addition, with the growth of
the scale of the data sets, the classical sample selection algorithm has been insufficient to cope
with large datasets, so the high efficient sample selection algorithm es a research
hotspot in recent years.
In order to deal with problems of large datasets classification, this paper proposed
sample selection algorithms which can be used for differe