基于数据降维的机器学习分类应用研究

发布时间：2018-01-26 00:08

本文关键词： 数据降维机器学习分类问题手写数字识别　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：在当今社会,信息技术的进步使得数据获取成本不断降低,随着海量数据的不断涌现,数据的维度也不断攀升。一般来说数据的维度越高,计算复杂度越大,数据中的噪声与冗余特征产生的负面影响越严重。因此,如何降低数据的维度,提高数据分类精度,已经成为机器学习领域的重要问题。本文围绕数据降维对机器学习分类效果所产生的影响这一问题展开研究。首先构建降维分类的分析架构,利用非线性降维局部线性嵌入(LLE)和线性降维主成分分析(PCA)这2种不同的降维方法与梯度提升决策树(GBDT)、随机森林(Random Forest)、支持向量机(SVM)、K近邻(KNN)、Logistic回归分类(Logistic Regression)共5种机器学习分类方法进行了组合。然后利用手写数字识别数据集,对这5种分类方法在不同降维方法下所产生的不同维度数据集上的分类表现进行了分析。分析表明,利用合适的降维方法进行降维分类可以有效的提升分类的准确率;非线性降维方法的降维分类效果普遍好于线性降维;不同机器学习分类算法对于维度的敏感度有着显著的差异;降维分类可以在保持分类精度的同时大大减少模型训练的时间。
[Abstract]:In today's society, with the development of information technology, the cost of data acquisition is continuously reduced. With the continuous emergence of massive data, the dimension of data is also rising. Generally speaking, the higher the dimension of data, the greater the computational complexity. The negative effects of noise and redundant features in data are more serious. Therefore, how to reduce the dimension of data and improve the accuracy of data classification. It has become an important problem in the field of machine learning. This paper focuses on the impact of data dimensionality reduction on the classification effect of machine learning. Firstly, the analysis framework of dimensionality reduction classification is constructed. Two different dimensionality reduction methods and gradient lifting decision tree (GBDTT) are proposed by using nonlinear dimensionally reduced local linear embedding (LLEE) and linear dimensionally reduced principal component analysis (PCAA). Random Forestland, support Vector Machine (SVM). Five machine learning classification methods were combined by Logistic regression classification and Logistic regression. Then handwritten numerals were used to recognize the data sets. The classification performance of these five classification methods on different dimension data sets produced by different dimensionality reduction methods is analyzed. Using proper dimensionality reduction method can effectively improve the accuracy of classification. The effect of nonlinear dimensionality reduction method is generally better than that of linear dimensionality reduction method. There are significant differences in the sensitivity of different machine learning classification algorithms to the dimensions. Dimension reduction classification can greatly reduce the time of model training while maintaining classification accuracy.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】