基于互信息的混合属性数据特征选择方法

Mutual information based feature selection for mixed attributes data

对于混合属性条件下的特征选择问题,给出了一种基于互信息的特征选择方法。首先,将互信息的定义推广到混合属性,在给出其计算方法的基础上,利用互信息定义了一种新的混合属性间的相关性度量;然后,通过对过滤式特征选择中的评价准则进行改造,完成原始特征的初选;最后,以估算精度为标准,对过滤式特征选择中的参数进行优化,确定最终的特征子集。实验结果表明:该方法具有较好的稳定性和估算精度。

For feature selection with mixed attributes data, a mutual information based method is pro- posed. Firstly, the concept of mutual information is extended to mixed attributes. By presenting a method for calculating mutual information between continuous and discrete attributes, a relevance measurement between mixed attributes is defined; Secondly, the features are evaluated by reconstructing the evaluating criterion in filter feature selection; Finally, features are selected by optimizing the parameter in filter feature selection with estimation accuracy criterion. Experimental results show that the method acquires preferable stability and estimation accuracy;