基于PSO面向K近邻分类的特征权重学习算法  

PSO Based Feature Weighting Algorithm for KNN

作　　者：任江涛[1] 卓晓岚[1] 许盛灿[1] 印鉴[1] 

机构地区：[1]中山大学计算机科学系,广州510275

出　　处：《计算机科学》2007年第5期187-189,共3页Computer Science

基　　金：国家自然科学基金资助(60573097);广东省自然科学基金资助(04300462;05200302)

摘　　要：特征权重学习是基于特征赋权的K近邻算法需要解决的重要问题之一,传统上提出了许多启发式的学习方法。近年来,随着进化计算技术在模式识别及数据挖掘领域的广泛应用,基于进化计算的权重学习和距离学习方法也得到越来越多的重视。本研究针对基于特征赋权的K近邻算法的权重学习问题,提出了一种基于PSO进行权重学习的算法PSOKNN,通过与传统KNN、GAKNN及ReliefKNN的实验比较分析表明,该方法可有效地搜索出合适的特征权重,获得较好的分类精度并淘汰冗余或无关的特征。Feature weighting is one of the important problems for feature weighting based KNN algorithm, and many heuristic methods have been employed to solve the problem traditionally. Recently, with the wide applications of evolutionary computation in pattern recognition and data mining areas, the evolutionary computation based feature weighting and distance learning method have got more and more attentiorh According to the feature-weighting problem of KNN, the paper proposes a PSO based feature weighting algorithm named PSOKNN. In comparison with other rdethods such as traditional KNN, GAKNN and ReliefKNN, the experiments show that the PSOKNN can get suitable feature weights, result in good classification accuracy and get rid of more redundant or irrelevant features.

关 键 词：特征赋权 K近邻分类 粒子群算法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]