多标签特征选择的两种算法研究

发布时间：2018-01-15 19:17

本文关键词：多标签特征选择的两种算法研究　出处：《闽南师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：区别于传统的监督学习中每个样本只属于一个类别标签的学习框架,多标签学习中的一个样本同时标注为多个类别标签的学习框架能够更有效地分析现实世界中所存在的问题。多标签学习是一个具有挑战性的研究课题,来源于文本分类中所遇到的多义性问题。近十几年来,吸引了越来越多的研究人员的兴趣,并应用到很多新的领域,如生物信息学、多媒体自动标注以及情感分类等。学者们已经提出了一系列多标签分类算法,然而高维数据中的大量不相关、冗余的特征导致了分类器性能的降低。因此,特征选择作为一种高效的数据降维技术,在多标签学习中可以起到关键的作用。目前,针对多标签学习的特征选择问题的研究相对较少。为此,本文围绕该问题,将非负矩阵分解和流形学习技术应用到多标签学习中,提出了两种多标签特征选择算法。主要研究工作如下:1.提出了一种基于非负稀疏表示的多标签特征选择算法。首先,我们引入子空间学习用于多标签特征选择,并在矩阵分解过程中对指示矩阵加以非负和稀疏约束。然后,我们设计一种高效的矩阵更新迭代算法,用以求解矩阵的非负约束和2,1L-范数最小优化的融合问题。最后,实验结果表明了该算法在多标签数据集上有着较优的特征选择性能。2.提出了一种基于特征流形和稀疏正则化的多标签特征选择算法。首先,我们利用最小二乘回归模型把多标签特征选择形式化成一个带有正则项的矩阵分解问题,用回归系数矩阵评估特征重要度。然后,我们将特征流形和稀疏约束嵌入到优化问题框架中,试图获得平滑的、行稀疏的回归系数矩阵。最后,我们设计出一种高效的矩阵更新迭代算法来求解优化问题,并用实验验证了所提出的算法的有效性。
[Abstract]:Different from the traditional supervised learning, each sample belongs to a learning framework with only one category label. A learning framework in which a sample is labeled as a multi-class label in multi-label learning can more effectively analyze the problems in the real world. Multi-label learning is a challenging research topic. In recent years, it has attracted more and more researchers' interest and applied to many new fields, such as bioinformatics. Researchers have proposed a series of multi-label classification algorithms, but a large number of irrelevant data in high-dimensional data, redundant features lead to the performance of classifier degradation. Feature selection, as an efficient data dimensionality reduction technique, can play a key role in multi-label learning. At present, there are few researches on feature selection for multi-label learning. In this paper, the nonnegative matrix decomposition and manifold learning techniques are applied to multi-label learning. Two multi-label feature selection algorithms are proposed. The main research work is as follows: 1. A multi-label feature selection algorithm based on non-negative sparse representation is proposed. We introduce subspace learning for multi-label feature selection and apply nonnegative and sparse constraints to the indicator matrix during matrix decomposition. Then we design an efficient matrix update iterative algorithm. It is used to solve the fusion problem of nonnegative constraints of matrix and minimum optimization of 2L L norm. Finally. Experimental results show that the algorithm has better feature selection performance on multi-label datasets. 2. A multi-label feature selection algorithm based on feature manifold and sparse regularization is proposed. We formalize multi-label feature selection into a matrix decomposition problem with regular terms by using least square regression model and evaluate feature importance by regression coefficient matrix. We embed eigenmanifold and sparse constraints into the framework of the optimization problem, trying to obtain a smooth, row sparse regression coefficient matrix. Finally. An efficient matrix update iterative algorithm is designed to solve the optimization problem. The effectiveness of the proposed algorithm is verified by experiments.
【学位授予单位】：闽南师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】