Abstract:For attribute reduction of neighborhood decision system,in the view of problem that it can′t reflect neighborhood information granules decision information accurately under principle of minimized error rate of neighborhood decision,with sample′s neighborhood information and its decision distribution being combined,a correlation measure which can reflect relation of condition attribute subset and decision attribute is proposed.Firstly,neighborhood relation matrix is introduced to convert set operation into matrix operation for neighborhood relation.Secondly,monotonicity of granularity of the new metric is proved,with combination of sort idea and symmetry of neighborhood matrix,an improved Single Attribute Neighborhood Relationship Matrix algorithm(SANRM)is prompted and a heuristic Neighborhood Relation Matrix Attribute Reduction algorithm(NRMAR)is constructed.Finally,experiments of UCI datasets show that NRMAR can effectively select attributes and maintain or improve classification ability of the datasets simultaneously.