Accelerator for multi-granularity attribute reduction

By considering the information granulation in Granular Computing, the concept of the multi-granularity is important. It is mainly because different results of information granulation will imply different levels of granularity. Nevertheless, multi-granularity has been paid less attention to the problem of attribute reduction in rough set which is regarded as one of the most important mathematical tools in Granular Computing. Therefore, how to search the multi-granularity reduct will be mainly explored in this paper. Different from the previous studies which generate reduct by using one and only one granularity, multi-granularity reduct is actually a set of the reducts derived from multiple levels of granularity. A natural way for computing multi-granularity reduct is to repeat the process of searching reduct for each level of granularity. Obviously, such an approach is time-consuming. To fill such a gap, an acceleration strategy is introduced into the process of searching multi-granularity reduct. Our acceleration strategy can be respectively realized through considering two variations of granularity: 1) from a finer granularity to a coarser granularity; 2) from a coarser granularity to a finer granularity. Such two variations indicate that the reduct related to previous granularity may have guidance on the computation of reduct related to the present granularity. Consequently, two accelerators are designed for speeding up the process of finding multi-granularity reduct. The experimental results over 16 UCI data sets show that our accelerators can not only reduce the elapsed time of searching attributes significantly, but also select attributes which will not contribute to a poorer classification performance. This study suggests new trends concerning the problem of attribute reduction and the corresponding searching strategy.