Granular ball guided selector for attribute reduction

摘要

In this study, a granular ball based selector was developed for reducing the dimensions of data from the perspective of attribute reduction. The granular ball theory offers a data-adaptive strategy for realizing information granulation process. It follows that the obtained granular balls can be regarded as the fundamental units of sampling and thereafter, the procedure of deriving the reduct(s) can be redesigned from a novel perspective. Firstly, the set of all granular balls is sorted based on their purities, following which each granular ball is considered as a group of samples, this is actually a process of sampling. Secondly, a potential reduct is derived over the first granular ball. Thereafter, a reduct over the subsequent granular ball can be obtained through correcting this potential reduct. Repeat this process until the reduct over the last granular ball is generated. Finally, the last reduct will be further corrected for deriving the final result over the whole universe. By considering both the efficiency of searching the reduct(s) and the effectiveness of the obtained reduct(s), comprehensive experiments over a total of 20 UCI datasets clearly validated the superiority of our approach against six well-established algorithms.