项目名称: 基于数据集“粒结构”和几何结构的子空间学习算法研究
项目编号: No.61203240
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 自动化学科
项目作者: 魏莱
作者单位: 上海海事大学
项目金额: 24万元
中文摘要: 子空间学习是应用最为广泛的特征提取方法之一。在子空间学习算法中,数据集蕴含的几何结构信息已被较深入的研究和利用,但对数据集的"粒结构"信息仍然缺乏足够的分析与讨论。本项目旨在将数据集的"粒结构"信息与几何结构信息相结合,将这两种数据集的结构信息综合融入子空间学习算法的设计中:结合数据集几何结构发现数据集的"粒结构",并设计数据点间的新的相似度及"数据粒"之间的测度,由此开发能够利用数据集两种结构信息的子空间学习算法;分析不同的子空间算法,根据算法特性及数据集"粒结构"具有层次性的特点,开发能够利用不同层次"粒结构"信息的集成子空间学习算法,并提出一个基于数据集"粒结构"的子空间学习算法框架;结合子空间学习算法的发展方向,将数据集"粒结构"信息引入到子空间算法发展的最新成果,发展更高效、鲁棒的子空间学习算法。本项目研究将为特征提取方法研究提供技术方法和理论指导,具有重要的理论及应用价值。
中文关键词: 子空间学习;粒结构;几何结构;稀疏表示;低秩表示
英文摘要: Subspace learning is one of most popular algorithms for feature extraction. In the studies of subspace learning, the geometric structures of data sets have already been made an in-depth exploration, but few attentions are paid on the "granular structures" of the data sets. This project aims to research the subspace learning algorithms with the geometric structure information and the "granular structure" information of data sets. Firstly, we will propose the methods for finding the "granular structure" of data sets with the help of the geometric structure information. The new similarity definition of the data samples and the metrics of "data granulas" will be proposed in this procedure. Based on these definitions, we hope to develop more reasonable subspace learning algorithms which are able to use both the two kinds of structure information of data sets. Secondly, we will analysis the existing subspace learning algorithms. According to characters of the different subspace learning algorithms and the hierarchical "granular structure" of data sets, we aim to propose the ensemble subspace learning algorithms which are capable to use the multi-level "granular structure" infoimation of the data sets. Moreover, we expect to form a unified framework for subspace learning algorithms which use the "granular structure" o
英文关键词: subspace learning;granularity structure;geometric structure;sparse representation;low-rank representation