摘要
针对微阵列数据样本量少、维度高的特点,结合当前数据降维方法中没有考虑特征与特征之间相关性的缺点,提出一种核最小二乘的特征基因选择方法。将解释变量空间通过非线性映射转换到高维空间上,再在高维空间上进行最小二乘回归,并采用极限学习机进行训练和预测。结果表明:对三种经典数据集的分类精度分别达到90.47%,88.89%,88.23%,高于传统的机器学习算法,充分表明本方法的优越性。
As quantity of microarray data sample is little and dimension of each sample is high, combined with disadvantages that in current data dimension reduction methods, correlation between features is not considered, put forward a kind of kernel-based least squares method for feature gene selection. Map explaining variable space to high dimension space, via nonlinear mapping transformation, and then carry out least-squares regression in high dimensional space; use extreme learning machine for training and predicting. The results show that c]assification precision of the three kinds of classic data set is 90.47 %, 88.89 % , 88.23 %, which is higher than traditional machine learning algorithms, which fully demonstrates superiority of this method.
出处
《传感器与微系统》
CSCD
2016年第5期146-148,153,共4页
Transducer and Microsystem Technologies
关键词
微阵列分类
基因选择
核最小二乘
极限学习机
microarray classification
gene selection
kernel least squares
extreme learning machine
