NOTE Communicated by Geoffrey Goodhill Improved Multidimensional Scaling Analysis Using Neu.pdfVIP

NOTE Communicated by Geoffrey Goodhill Improved Multidimensional Scaling Analysis Using Neural Networks with Distance-Error Backpropagation Llu´ıs Garrido ` Departament d’Estructura i Constituents de la MateriaIFAE, Universitat de Barcelona, E-08028 Barcelona, Spain Sergio G´omez ` Departament d’Enginyeria Informatica, Universitat Rovira i Virgili, E-43006 Tarrag- ona, Spain Jaume Roca ` Departament d’Estructura i Constituents de la MateriaIFAE, Universitat de Barcelona, E-08028 Barcelona, Spain We show that neural networks, with a suitable error function for back- propagation, can be successfully used for metric multidimensional scal- ing (MDS) (i.e., dimensional reduction while trying to preserve the origi- nal distances between patterns) and are in fact able to outdo the standard algebraic approach to MDS, known as classical scaling. 1 Introduction A standard problem in multidimensional scaling analysis is to map a col- lection of patterns, represented as points in an n-dimensional space xa n a 1 p to a lower-dimensional space in such a way that the distances between the projected points resemble as closely as possible the distances between the original ones. More precisely, given the collection xa , with Euclidean distances be- tween pairs a b of patterns: n 2 d xa x ab b one has to find a map, : n m with m n, such that it minimizes the quadratic distance-error function E 1 dn dm2 2 ab ab ab c