%A JIANG Ze-hua, WANG Yi-bo, XU Gang, YANG Xi-bei, WANG Ping-xin
%T Multi-scale Based Accelerator for Attribute Reduction
%0 Journal Article
%D 2019
%J Computer Science
%R 10.11896/jsjkx.181102031
%P 250-256
%V 46
%N 12
%U {https://www.jsjkx.com/CN/abstract/article_18801.shtml}
%8 2019-12-15
%X The neighborhood rough set measures the similarity between samples by radius,consequently,different radii naturally construct the rough approximations with different scales.Traditional attribute reduction based on neighborhood rough set is frequently executed over multi-radius.The aims are to select an attribute subset with better generalization performance or to explore the trend line of the performances of the reducts in terms of different scales.However,it should be emphasized that the process should be repeatedly executed for each scale,if the traditional algorithm based on heuristic searching is used to compute the multiple scale reducts.The time consumption of computing reducts is too high to be accepted,especially the number of the scale is more,the time consumption will be more.To solve such a problem,the multi-scale based accelerated strategy for attribute reduction was proposed by means of the changing of radius.This strategy considers two trends of changing radius,from smaller radius to greater radius and from greater radius to smaller radius.Moreover,the traversal size of samples and attributes is reduced.Therefore,the current process to find reduct is executed based on the reduct related to the previous radius,which follows that the forward or backward heuristic searching for adding and deleting attributes can be realized.To validate the effectiveness of the accelerated strategy,8 UCI data sets,10-fold cross-validation and 20 radii were employed to conduct the experiment,and the time consumption of computing different reducts and the classification of different reducts were compared.The experimental results over 8 UCI data sets show that the proposed forward or backward accelerated searching can significantly reduce the time consumptions of finding reducts if the case of multi-scale is considered.Moreover,the proposed approach will not significantly decrease the classification performance,and can reduce the degree of over-fitting.