Estimation of boolean factor analysis performance by informational gain

Abstract

To evaluate the soundness of multidimensional binary signal analysis based on Boolean factor analysis theory and mainly of its neural network implementation, proposed is a universal measure - informational gain. This measure is derived using classical informational theory results. Neural network based Boolean factor analysis method efficiency is demonstrated using this measure, both when applied to Bars Problem benchmark data and to real textual data. It is shown that when applied to the well defined Bars Problem data, Boolean factor analysis provides informational gain close to its maximum, i.e. the latent structure of the testing images data was revealed with the maximal accuracy. For scientific origin real textual data the informational gain provided by the method happened to be much higher comparing to that based on human experts proposal. © Springer-Verlag Berlin Heidelberg 2010.