Abstract: The k-NN rules and their modifications offer usuallyvery good performance. The main disadvantage of thek-NN rules is the necessity of keeping the referenceset (i.e. training set) in the computer memory.Numerous algorithms for the reference set reductionhave been already created. They concern the 1-NN ruleand are based on the consistency idea. The 1-NN ruleoperating with a consistent reduced set classifiescorrectly, by virtue of consistency, all objects fromthe original reference set. Quite different approach,based on partitioning of the reference set into somesubsets, was proposed earlier by the present authors.The gravity centers of the subsets form the reducedreference set. The paper compares the effectiveness ofthe two approaches mentioned above. Ten experimentswith real data concerning remote sensing data arepresented to show the superiority of the approach basedon the reference set partitioning idea. !8
展开▼
