基于改进Shapley权力指数的特征选择算法

Probabilistic Model and Improved Shapley Power Index Based General Feature Selection Algorithm

针对特征选择算法对不同类型的数据集性能不稳定的问题,提出一种基于概率模型与改进Shapley权力指数的通用特征选择算法.首先,计算特征对类簇表征与类簇判别的重要性值;然后,计算特征对类簇的不确定度;最终,合并特征的重要性值与不确定度,提取合适的特征.因为概率模型对数据类型、数据缺陷具有较好的鲁棒性,所以对不同的数据集获得了稳定、高性能的特征选择效果.基于人工合成数据与benchmark数据集的实验结果表明,本算法对不同的数据集保持了稳定的特征选择效果,优于其他算法.

Aimed at the problem that feature selection algorithms show unstable performance to different types of datasets,aprobabilistic model and improved Shapley power index based general feature selection algorithm.Firstly,the importance values of features to class representation and discrimination are computed.Secondly,the uncertainness values of features to classes are computed.Lastly,the importance values and uncertainness values are merged to abstract suitable feature.Because probabilistic model performs good robustness to data types and data imperfection,stable feature selection results to different datasets are got with high performance.Synthetic datasets and benchmark datasets based experiments results show that the proposed algorithm show stable feature selection effect to different datasets and is better than the other algorithms.