Multiclass classification based on binary classifiers: On coding matrix design, reliability and maximum number of classes
Slava Voloshynovskiy, Oleksiy Koval, Fokko Beekhof, Taras Holotyak · 2009
In this paper, we consider the multiclass classification problem based on independent set of binary classifiers. Each binary classifier represents the output of quantized projection of training data onto a randomly generated orthonormal basis vector thus producing a binary label. The ensemble of all binary labels forms an analogue of a coding matrix. The properties of such kind of matrices and their impact on the maximum number of uniquely distinguishable classes are analyzed in this paper from an information theoretic point of view. We also consider a concept of reliability for such kind of coding matrix generation that can be an alternative way for other adaptive training techniques and investigate the impact on the bit error probability. We demonstrate that it is equivalent to the considered random coding matrix without any bit reliability information in terms of recognition rate.