Template-type: ReDif-Paper 1.0
Author-Name: Carrizosa,Emilio 
Author-Name: Martín-Barragán,Belén
Author-Name: Plastria,Frank 
Author-Name: Romero Morales,Dolores 
Author-workplace-name: METEOR
Title: A dissimilarity-based approach for Classification 
Abstract: The Nearest Neighbor classifier has shown to be a powerful tool for multiclass classification. In this note we explore both theoretical properties and empirical behavior of a variant of such method, in which the Nearest Neighbor rule is applied after selecting a set of so-called prototypes, whose cardinality is fixed in advance, by minimizing the empirical mis-classification cost. With this we alleviate the two serious drawbacks of the Nearest Neighbor method: high storage requirements and time-consuming queries. The problem is shown to be NP-Hard. Mixed Integer Programming (MIP) programs are formulated, theoretically compared and solved by a standard MIP solver for problem instances of small size. Large sized problem instances are solved by a metaheuristic yielding good classification rules in reasonable time. 
Keywords: operations research and management science; 
Series: Research Memoranda
Creation-Date: 2005
Number: 045
File-URL: http://arnop.unimaas.nl/show.cgi?fid=3753
File-Format: application/pdf
File-Size: 380505
Handle: RePEc:unm:umamet:2005045