A SMALL STORE PROBABILITY DENSITY FUNCTION ESTIMATOR FOR INCOMPLETE DATA CLASSIFICATION

In the paper the problem of designing a pattern recognition system for processing incomplete pattern vectors is considered. An efficient method of integrating the small core probability density function (p.d.f.) estimator employing Gaussian kernels with general parameter matrices has been proposed. As a result these general kernels satisfy the basic requirements of integrability and, therefore, they can be used in p.d.f. estimators for classification systems processing incomplete pattern vectors. In comparison with the Gaussian kernel having a diagonal parameter matrix, the general kernel is better suited for reconstructing multivariate p.d.f. of pattern vectors with correlated components. Also determination of the optimal parameters of the general kernel is much simpler for the use of the computationally demanding maximum likelihood parameter estimation method can be obviated.