Noise Resilience on Supervised Classification of Shortest-Length and Coarsest-Granularity Reducts and Constructs

This study evaluates noise resilience in supervised classification using shortest-length and coarsest-granularity reducts and constructs under a common evaluation framework. We analyze original datasets and training sets distorted by random attribute-value noise across 20 real-world datasets and four supervised classifiers. Experimental results, validated with the Wilcoxon signed-rank test, indicate no statistically significant difference in classification accuracy between reducts and constructs across the evaluated noise levels. These findings suggest that, for the considered random-noise model, the choice between shortest-length and coarsest-granularity subsets has limited impact on supervised classification accuracy.