Attribute reduction is one of the important research points in rough set theory. The goal of β distribution reduction in interval-valued decision systems is to keep the corresponding β distribution of objects unchanged. In actual needs, attribute reduction usually only needs to focus on specific decision class rather than all decision classes. This paper proposes a theoretical of class-specific β distribution reduction in interval-valued decision systems. First of all, the basic concept of class-specific β distribution reduction is defined, and then the discernibility matrix corresponding to the class-specific β distribution is constructed. Finally, a class-specific β distribution reduction algorithm based on the discernibility matrices is proposed. In experiments, six UCI data sets are used to compare reduction results and reduction efficiency of BRADM algorithm and CSBRADM algorithm. The experiments results show that the reduction results of class-specific algorithm can keep the β distribution for class-specific unchanged, and reduction length of algorithm for specific class is less than or equal to reduction length of algorithm for all classes, and the CSBRADM algorithm efficiency is higher than the BRADM algorithm.