采用特征相关性差异优化距离的改进k近邻算法.pdfVIP

3178 2011 , Vo1.32, No.9 计算机工程与设计 Computer Engineering and Design 采用特征相关性差异优化距离的改进 k 近邻算法 周靖，刘晋胜 (广东石油化工学院计算机与电子信息学院，广东茂名 525000) 摘 要:为有效地解决 k 近邻算法训练样本规模及分类精度间的矛盾，提出了一种采用特征相关性差异优化距离的改进算 法 (FCDKNN) 。该算法将特征煽值与其分布概率的来积作为特征相关性的概念，在此基础上定义围绕特征相关性差异的样本 距离测度，明确特征在类别上的重要性及相关性，在小样本情况，下提取针对分类的大量有效信息，以增强算法的全局优化 能力.对比仿真实验结果表明，该算法在保持效率的情况下分类精度得到了???大地提高. 关键词 :k 近邻;特征;相关性;差异;距离 中国法分类号: TP301 文献标识码 :A 文章编号: 1000-7024 (201 1) 09-3178-04 Improved algorithm of adopting feature correlation ditference to optimize distance ZHOU Jing, LIU Jin-sheng (College of Computer and Electronic Information, Guangdong University of Pe位ochemical Technology, Maoming 525000, China) Abstract: To solve the contradiction of k-nearest neighbor between training sarnples size and c1assification accuracy e 能ctive 以 anim- proved algorithm (feature correlation di 仔erence k-nearest neighbor, FCDKNN) adopting correlation difference to optimize distance is proposed, the impacted algorithm takes the product between the entropy of feature and distribution probability of feature as the concept of correlation. and a distance measure of sarnples is defined around feature correlation difference. the important and the correlation of feature on c1assification are c1eared. a lot of useful information about classification is extracted in small sarnples for employing global optimization ability. The con往ast simulation experiment shows that the improved algorithrn raised the rate of c1assification acc四百.cy greatly. meanwhile it also maintains efficiency of c1assification. Key words: k-nearest neighbor; feature; cor