项目名称: 特征选择中的全局最优搜索策略研究
项目编号: No.61202134
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 严慧
作者单位: 南京理工大学
项目金额: 24万元
中文摘要: 特征选择是一门多学科的交叉研究方向,它涉及统计学、数据挖掘、模式识别和机器学习等相关学科,在生物信息学、医学、信息检索等诸多领域具有广泛应用前景。传统的特征选择以代价换取简单、快速的搜索策略实现,不能保证最优,甚至有时获得很不理想的特征子集。本项目以高维空间中的数据为研究对象,以全局最优特征子集的搜索方式为科学问题,深入挖掘特征抽取中的投影方式与特征选择中最优搜索之间的紧密联系,实现了两者在特定条件下的可通行。本项目研究的预期成果是:(1)采用特殊的矩阵投影方式替代传统特征选择流程中的逐渐增加(或删除)特征的搜索路径,力图构造一个存在全局最优解的组合优化问题;(2)定义出与矩阵投影方式相匹配的特征评估准则,包括基于信息熵、稀疏表示理论、图论的度量方式;(3)设计针对0-1大规模稀疏矩阵求解的优化算法。本项目对拓展传统特征搜索理论和算法有十分重要的意义,且具有十分广阔的应用前景。
中文关键词: 特征选择;同步特征选择;最优性;;
英文摘要: Feature selection is a multi-knowledge crossed research direction, since It refers to statistics, data mining, pattern recognition, machine learning and so on. Feature selection has been widely applied to Bioinformatics, medical science, information retrieval and so on. Traditional feature selection is achieved as a simple and fast search strategy, that can not guarantee an optimal solution and sometimes the obtained feature subset is far from perfect. This project, studying data in the high-dimensional space, solves the search solution for the global optimal feature subset. This project explores the relationship between matrix projection in feature extraction and optimal search direction and further proves the equivalence of both under some specified conditions. The anticipated achievements of this project are: (1) Instead of adding (or deleting) one feature gradually in traditional feature selection, this project adopts a special matrix projection. And it tries to construct a combinational optimization problem which has an global optimal solution.(2) This project defines some feature evaluation criterion that match the matrix projection, including evaluation criterions based on information entropy, sparse representation theory, subspace learning.(3) This project designs optimization algorithm for 0-1 large-sc
英文关键词: feature selection;joint feature selection;optimality;;