基于非负矩阵分解的代价敏感特征选择

Cost-Sensitive Feature Selection based on Non-Negative Matrix Factorization

数据降维可降低分析处理多维数据的复杂度和成本.特征选择是常见的数据降维方法.传统的特征选择算法更多关注算法的分类性能,忽略了对选择过程中产生的测试代价(Cost-test)的考虑.基于此提出一种新的基于非负分解的代价敏感特征选择方法(NmfCt).NmfCt算法构造的目标函数能够同时约束重建误差最小和测试代价最小,在对数据进行预处理降维的同时,不但能确保较好的分类正确率(Accuracy),而且还能保持较低的测试代价.

Data dimension reduction can lower the complexity and cost of the analysis of multi-dimensional data, and feature selection is an effective method. Traditional feature selection methods tend to consider the classification performance of algorithm, meanwhile, ignore the cost-test in selecting. Therefore, according to the non-negative matrix factorization, a new cost sensitive feature selection method, namely NmfCt algorithm, is proposed. The ob- jective function constructed by NmfCt algorithm can let the reconstruction error in minimum and have a minimum testing Cost at the same time. NmfCt algorithm can not only ensure classification accuracy in data dimension reduc- tion, but also maintain the testing at low cost.