Abstract: Feature gene for classification is one of important problems in gene expression data analysis. Current methods ignore that gene expression is unbalanced in different classes. The paper introduces a new feature selection method for any sample. The method presents a new heuristic algorithm that is composed of an improved difference between classes and an original undulation inside the class. The experimental results show that the method is effective on selecting feature genes for unbalanced multi-class sample and advancing classification capability of classifiers.

Key words: feature selection, multi-class, classifier, gene expression profile

摘要： 分类特征基因是基因表达谱数据分析中的重点，目前的特征基因选择方法均没有考虑到基因在不同类别中分布失衡给特征基因选择算法带来的影响。提出一种样本无关的特征基因选择方法，该方法利用改进地类间差异函数和类内波动函数，根据两个函数的一致性选择每个类别的鉴别基因。该方法不仅适用于多类样本，对于各类样本数量不均衡以及基因在各类中分布失调的样本同样有效。实验结果表明，该方法确保了特征矢量的均衡性，提高了分类器的分类性能。

关键词: 特征选择, 多类, 分类器, 基因表达谱