Neighborhood attribute reduction: a multi-criterion approach

Jingzheng Li, Xibei Yang, Xiaoning Song, Jinhai Li, Pingxin Wang, Dong-Jun Yu

Abstract: Though attribute reduction defined by neighborhood decision error rate can improve the classification performance of neighborhood classifier via deleting redundant attributes, such reduction does not take the variations of classification results into account. To fill this gap, a multi-criterion based attribute reduction is proposed, which considers both neighborhood decision error rate and neighborhood decision consistency. The neighborhood decision consistency is used to measure the variations of classification results if attributes change. Following the novel attribute reduction, a heuristic algorithm is also designed to derive reduct which aims to obtain less error rate and higher consistency simultaneously. The experimental results on 10 UCI data sets show that the multi-criterion based reduction can not only improve the decision consistencies without decreasing the classification accuracies significantly, but also bring us more stable reducts. This study suggests new trends concerning criteria and constraints in attribute reduction.