SCHEMES OF COMBINING DISCRIMINANT FUNCTIONS FOR INCREASING THE CLASSIFICATION ACCURACY IN AN ENSEMBLE OF DATA SOURCES
M.M. Lange, S.V. Paramonov
a:2:{s:4:"TEXT";s:99:"Federal Research Center “Computer Science and Control”, Russian Academy of Sciences, Moscow, Russia";s:4:"TYPE";s:4:"text";}
Keywords: classification, ensemble of sources, fusion scheme, error probability, mutual information, Hamming distortion measure, rate-distortion function, discriminant function, entropy, redundancy
Abstract
Given an ensemble of datasets, we study the object classification accuracy in terms of the error probability depending on the amount of processed information using various fusion schemes. Schemes of combining weak discriminant functions in each dataset as well as in an ensemble of different modality datasets are suggested. For the proposed fusion schemes, the redundancy of the error probability relative to the information-theoretic lower bound defined by the modified rate-distortion function with the Hamming distortion measure is evaluated. The experimental evaluations in datasets of signature and face images show a decrease in the error probability and its redundancy with the amount of the processed information being increased by combining weak discriminant functions.
|
