Abstract: Support Vector Machine—Recursive Feature Elimination（SVM-RFE） is one of state-of-the-art method for gene selection.SVM-RFE was originally designed to solve binary feature selection problems and has been extended to solve multiclass problems in several recent studies.This paper illustrates the limitations of the present multi-class gene selection methods from the perspective of Pareto Optimum，describes a new procedure for selecting significant genes for each class，and proposes a new implementation for SVM-RFE. Experiments on 8 cancer and tumor gene expression dataset demonstrate its superiority over two other RFE methods.By considering each class during the gene selection stages，the new method can identify genes leading to more accurate classification.

Key words: gene expression data, multi-classification, gene selection, support vector machine

摘要： 基于支持向量机的递归特征消除（SVM-RFE）是目前最主流的基因选择方法之一，是为二分类问题设计的，对于多分类问题必须要进行扩展。从帕累托最优（Pareto Optimum）的概念出发，阐明了常用的基因选择方法在多分类问题中的局限性，提出了基于类别的基因选择过程，并据此提出一种新的SVM-RFE设计方法。8个癌症和肿瘤基因表达谱数据上的实验结果证明了新方法优于另两种递归特征消除方法，为每一类单独寻找最优基因，能够得到更高的分类准确率。

关键词: 基因表达谱, 多分类问题, 基因选择, 支持向量机