Abstract: For the issue that the traditional feature selection method in determining redundant parameter β based on the Mutual Information（MI）, a kind of improved feature selection algorithm for NMIFS-FS2 is proposed. This algorithm input is characterized by a combination of MI and between classes, instead of the traditional algorithms in a single feature MI and between classes when selecting continuous or discrete features, solving the problem that redundancy parameter β is very difficult to determine, and expands the scope of application. Two sets of experiments conducted to verify the validity of this algorithm. Experimental results show that this algorithm, compared to several traditional classification algorithms, has better robustness, stability and efficiency.

Key words: input feature selection, optimal feature set, regular mutual information, classification, robustness

摘要： 针对基于互信息（MI）传统特征选择方法中要求确定冗余度参数[β]的问题，提出一种改进型特征选择算法NMIFS-FS2。该算法在对连续或离散特征进行选择时，输入为特征组合与类之间的MI，代替传统算法中单一特征与类之间的MI，解决了冗余度参数[β]很难确定的问题，扩大了应用范围。进行的两组实验验证了该算法的有效性。实验结果表明，相比几种传统的分类算法，该算法具有更好的鲁棒性、稳定性和高效性。

关键词: 输入特征选择, 最优特性集, 正则化互信息, 分类, 鲁棒性