联合不相关回归和非负谱分析的无监督特征选择

PDF下载HTML

Title:

Joint uncorrelated regression and non-negative spectral analysis for unsupervised feature selection

作者:

朱星宇¹, 陈秀宏²

1. 江南大学 人工智能与计算机学院，江苏 无锡 214122;
2. 江南大学 江苏省媒体设计与软件技术重点实验室，江苏 无锡 214122

Author(s):

ZHU Xingyu¹, CHEN Xiuhong²

1. School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China;
2. Jiangsu Key Laboratory of Media Design and Software Technology, Jiangnan University, Wuxi 214122, China

关键词:

不相关回归; 非负谱分析; 冗余特征; 局部结构学习; 无监督学习; 自适应图; 特征选择; 判别性特征

Keywords:

uncorrelated regression; non-negative spectral analysis; redundant features; local structure learning; unsupervised learning; adaptive graph; feature selection; discriminant feature

分类号:

TP391.4

DOI:

10.11992/tis.202012033

摘要:

在无标签高维数据普遍存在的数据挖掘和模式识别任务中，无监督特征选择是必不可少的预处理步骤。然而现有的大多数特征选择方法忽略了数据特征之间的相关性，选择出具有高冗余、低判别性的特征。本文提出一种基于联合不相关回归和非负谱分析的无监督特征选择方法(joint uncorrelated regression and nonnegative spectral analysis for unsupervised feature selection)，在选择不相关且具有判别性特征的同时，自适应动态确定数据之间的相似性关系，从而能获得更准确的数据结构和标签信息。而且，模型中广义不相关约束能够避免平凡解，所以此方法具有不相关回归和非负谱聚类两种特征选择方法的优点。本文还设计出一种求解模型的高效算法，并在多个数据集上进行了大量实验与分析，验证模型的优越性。

Abstract:

Unsupervised feature selection is an essential preprocessing step in the data mining and pattern recognition tasks of unlabeled high-dimensional data. However, most existing feature selection methods ignore the correlation between data features and select features with high redundancy and low discrimination. This paper proposes an unsupervised feature selection method based on joint uncorrelated regression and non-negative spectral analysis (Joint uncorrelated regression and nonnegative spectral analysis for unsupervised feature selection). It adaptively and dynamically determines the similarity relationship between data while selecting uncorrelated and discriminant features, so that more accurate data structure and label information can be obtained. Moreover, the generalized uncorrelated constraints in the model can avoid trivial solutions, so this method has the advantages of two feature selection methods: uncorrelated regression and non-negative spectral clustering. An efficient algorithm for solving the model is also designed, and a large number of experiments and analyses are carried out on multiple data sets to verify the superiority of the model.