Ensemble Feature Selection Based on Generalized Neighborhood Rough Model and Its Selective Integration
Guangming Wang · Xi'an Jiaotong Daxue xuebao · 2011
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.