Effects of data preprocessing on the efficiency of taxonomic neural networks

Objectives: For the classification of respondents drawn from different populations and described by different sets of variables, six taxonomic neural networks designed in a similar manner, but supplied with different preprocessors for input data transformation, were applied. Methods: The standard method used to classify objects was implemented by an Intruder neural network variation (Momirovic, Hosek. Popovic & Boli, 2002) codenamed Intruds. This network only standardizes input data, but does not perform any additional transformations over them. Results: Each of these networks consisted of Lebart`s adaptive perceptrons (Lebart, 1998) which classified objects by means of a backward error propagation technique at the output end. The initial classification in all the networks was carried out by transformation of a fuzzy classification, defined by the objects` position on the periphery of the hyperellipsoid formed by a semiorthogonal transformation of principal components of variables subjected to a monotonic transformation, into a hard classification. The final classification efficiency was checked by each network based on the outcome of a posteriori classification performed by Fisher`s linear classifiers in the total space of eventually transformed variables. The neural networks under the basic code name Invader perform nonlinear monotonic transformations over these data. Invader 1 transforms variables into sigmoid (0,1) form, Invader 2 transforms variables into sigmoid (-1,1) form, and Invader 3 – into staircase (-1, 0, 1) form. The neural network codenamed Bibltax reduces to zero all standardized values which differ from the expected value in the sample of objects by one half of the standardized deviation in order to eliminate noise produced by poorly differentiated objects; and the Exnetg neural network, a variation of the Exelnet neural network (Momirovic, 2002), transforms variables into standardized image form to eliminate noise produced by the unique variance of measurement instruments. Conclusions: Comparative analysis of the effectiveness of these taxonomic networks clearly shows that, in the all cases, better taxonomic effectiveness, or surer breaking up of the analyzed set of objects, is achieved by means of data preprocessing