Dynamic granularity selection based on local weighted accuracy and local likelihood ratio

Abstract

Granular computing aims to develop a granular view for interpreting and solving problems, in which granularity selection is a key problem and has received extensive attention in recent years. Existing studies select the same granularity for all samples. In fact, different samples may prefer to different granularities. To address this issue, dynamic granularity selection is proposed in this paper. Namely, granularity selection is considered with respect to specific sample. Two indices, denoted as local weighted accuracy and local likelihood ratio, are introduced to compute the weight of granularity. Subsequently, an algorithm called DGS−LWA-LLS is given for dynamic granularity selection, in which the granularity with the largest weight is considered to be optimal The weight of granularity is related to specific sample, thus the weights of a granularity may be different with different samples. Consequently, different granularities will be selected with respect to different samples. Experiments were carried out based on neighborhood granularity to explain the necessity of granularity selection and to validate the rationality and effectiveness of DGS−LWA-LLS.