广义邻域粗集下的集成特征选择及其选择性集成算法

Ensemble Feature Selection Based on Generalized Neighborhood Rough Model and Its Selective Integration

针对实际模式识别系统中样本特征常具有的连续值属性、高维性、强相关性和冗余性等影响分类效果的问题,在广义邻域粗集模型下提出一种集成特征选择及其选择性集成算法.该算法先提取样本特征并利用所提出的马氏距离分布熵评估其重要度,再基于特征重要度构建广义邻域粗集模型,并在此模型上以特征重要度为启发式信息设计基于蚁群算法的属性约简算法,然后通过改变广义邻域粗集模型参数的方式获得更多具有更大差异性的基分类器,最后利用主成分分析法对产生的基分类器进行选择性集成.模拟电路故障诊断结果表明,该算法比AdaBoost等算法取得的分类精度至少提高了2.6%.

A new ensemble feature selection method under the model of generalized neighborhood rough set is presented to improve the classification accuracy in actual pattern recognition systems. The importance degrees of sample features are evaluated by the distribution entropy of Mahalanobis distance （DEMD）, and the generalized neighborhood rough model is constructed based on resulting degrees. Then a fast attribute reduction algorithm that takes features importance degrees as heuristic information is designed to produce multiple reduction results for training basic classifiers. More basic classifiers with higher diversity are obtained through changing parameter values in the model, and these basic classifiers are then selectively integrated using the principal component analysis method. The fault diagnosis results of an analog circuit show that the proposed method increases classification accuracy by at least 2.6% compared with other methods such as Adaboost.