scientific article; zbMATH DE number 3999032

From MaRDI portal
Publication:4725538





Discussed is a problem of robustness of classification schemes formulated in the following way. Patterns coming from classes \(\omega_ 1,\omega_ 2,...,\omega_ c\) are viewed as elements of N-dimensional Euclidean space governed by probability density functions \(p_ i(x,\vartheta_ i)\), \(i=1,2,...,c\), with unknown parameters \(\vartheta_ i\). These parameters are estimated by means of plug-in rule (viz. on a basis of a sample of patterns \(A_ 1,A_ 2,...,A_ c\), \(card(A_ i)=n_ i\) coming from each class and chosen at random). In the problem discussed it is assumed that \(A_ 1,A_ 2,...,A_ c\) are noisy in the sense that \(A_ i\), \(i=1,2,...,c\) contains not only patterns of the i-th class itself but \(\epsilon_ i100 \%\) of patterns coming from the remaining classes. Two theorems are proved that describe the asymptotic behaviour of estimators of \(\vartheta_ i\) and a risk function with respect to small parameters \((\epsilon_ i,n_ i^{-1/2})\). Moreover a measure of robustness of classification rules is introduced and considered with respect to the above parameters.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4725538)