Boolean Factor Analysis by Attractor Neural Network

Abstract

Methods for the discovery of hidden structures of high-dimensional binary data rank among the most important challenges facing the community of machine learning researchers at present. There are many approaches in the literature that try to solve this hitherto rather ill-defined task. The Boolean factor analysis (BFA) studied in this work represents a hidden structure of binary data as Boolean superposition of binary factors complied with the BFA generative model of signals, and the criterion of optimality of BFA solution is given. In these terms, the BFA is a well-defined task completely analogous to linear factor analysis. The main contributions of the dissertation thesis are as follows: Firstly, an efficient BFA method, based on the original attractor neural network with increasing activity (ANNIA), which is subsequently improved through a combination with the expectation-maximization method(EM),so LANNIA method has been developed. Secondly, the characteristics of the ANNIA that are important for LANNIA and ANNIA methods functioning were analyzed. Then the functioning of both methods was validated on artificially generated data sets. Next, the method was applied to real-world data from different areas of science to demonstrate their contribution to this type of analysis. Finally, the BFA method was compared with related methods, including applicability analysis.