一种引入冗余控制的特征排序模型

中文摘要: 针对特征排序方法较少考虑特征之间的相关关系，导致选择的特征子集存在冗余的问题，提出一种引入冗余控制的特征排序模型。将特征子集判别能力最大且冗余程度最小作为模型的目标函数，以降低特征之间的冗余；使用贪心方法和非线性规划方法对模型进行求解。在9个开源数据上的实验及与特征排序方法比较表明，本模型在大部分数据上，所选择的特征子集能够获得更好的分类准确性且个数更少；使用非线性规划方法求解时，能够直接得到特征子集，有利于确定特征个数。本模型可用于特征之间存在冗余时的特征选择。
Abstract:Aimed at problems of feature redundancy caused by the fact that feature correlation was seldom considered in the feature ranking methods,a feature ranking model with redundancy control was proposed.Maximum discrimination ability and minimum redundancy of a feature subset were used as the objective functions of the very model so as to reduce the redundancy among features,and greed and non linear programming methods were employed to solve the model.Experiments were conducted on 9 public datasets and compared with feature ranking,and the result showed that the model can obtain a better classification accuracy and less feature size on most datasets.When non linear programming method is employed,the model can yield a feature subset,on benefit for determining the feature size.This model can be used when correlation exists among features.