考虑稳定性要求的特征选择方法

A feature selection algorithm taking account of stability

为了提高特征选择的稳定性和降低因样本数据变化引起的选择结果波动,提出了一种考虑稳定性要求的过滤式特征选择方法。不同于集成特征选择等现有的增强稳定性方法,该方法将特征的稳定性与相关性、冗余性一起作为特征评价准则,通过产生多个数据集来减少样本数据扰动,不断将新产生的选择结果迭代计算稳定性因子,并同时提高其在准则中的比重以使迭代收敛。最终将融合多次迭代信息的特征排序作为最终结果输出。实验表明,该方法能够在保持相当分类精度的基础上,能够较大幅度地提高选择结果的稳定性,达到兼顾分类精度与稳定性的目的。

To improve the stability of feature selection and reduce the fluctuations caused by the variation of sample data, a filter-type feature selection method considering the stability index is proposed. Unlike the integrated feature selec- tion and other methods, the propose method takes feature＇ s stability, together with the relevance and redundancy, as the evaluation criteria for feature selection, reduces the fluctuations of sample, data by producing multiple data sets, continuously puts new selection results into the iterative calculation of stability, and increases the proportion of the stability factor until the iteration is converged. At last, the achieved feature sequence fusing multi-iteration in- formation is taken as the final result of feature selection. The experimental results show that the proposed method can improve the stability of feature selection obviously, and reach the satisfied classification accuracy meanwhile.