Abstract: Attribute reduction is one of the important research topic in rough set theory.The minimal feature subset of a given information system can be obtained by attribute reduction.Classical attribute reduction in ordered decision system is about all decision classes in decision attribute.However,in practice,considering the preference of decision maker and the lack of some decision data,only attribute reduction of some specific decision classes is needed.Based on this consideration,in this paper,the dominance relations and lower approximation reduction in ordered decision table were reviewed,the single-class-specific and multi-class-specific lower approximation reduction in ordered decision table were presented,a discernibility matrix for this reduction was introduced,and an algorithm of lower approximation attribute reduction in ordered decision system was proposed.The multi-class-specific lower approximation reduction can degene-rate to single-class-specific reduction or classical reduction,so it is a more general reduction framework.In the experiment,6 sets of UCI data sets were used,3 single-class-specific reducts and 3 multi-class-specific reducts were calculated on each data set,and the reduction result and efficiency were compared with the result and efficiency of classical lower,upper approximation reduction and maximum distribution reduction.Experiment results show that when the number of selected specific classes is less than all decision classes,the reduct may be shorter,and the efficiency will be improved to varying degrees.

Key words: Attribute reduction, Discernibility matrix, Multi-class-specific, Ordered decision table, Rough set