Boolean Factor Analysis by the Expectation-Maximization Algorithm

0347803 - ÚI 2011 RIV DE eng C - Conference Paper (international conference)
Frolov, A. A. - Polyakov, P.Y. - Húsek, Dušan
Boolean Factor Analysis by the Expectation-Maximization Algorithm.
Proceedings of COMPSTAT 2010. Heidelberg: Physica Verlag, 2010 - (Lechevallier, Y.; Saporta, G.), s. 1039-1046. ISBN 978-3-7908-2603-6.
[COMPSTAT 2010. International Conference on Computational Statistics /19./. Paris (FR), 22.08.2010-27.08.2010]
R&D Projects: GA MŠMT(CZ) 1M0567; GA ČR GAP202/10/0262; GA ČR GA205/09/1079
Institutional research plan: CEZ:AV0Z10300504
Keywords : boolean factor analysis * generative model * information gain * efficiency measure
Subject RIV: IN - Informatics, Computer Science

Compared are e ciencies of two methods for Boolean factor analysis based on expectation-maximization technique. First one is Maximal Causes Analysis proposed by Lucke and Sahani (2008). Second one is Expectation-Maximization Boolean Factor Analysis, introduced here. Last method is strictly based on the general Boolean factor analysis generative model. Comparison is based on so called bars problem benchmark (Foldiak, 1990). Further informational theoretic measure of Boolean factor analysis e ciency is developed. Then it is shown that the e ciency of our Expectation-Maximization Boolean Factor Analysis method is higher then Maximal Causes Analysis in Boolean factor analysis model parameters entirety.
Permanent Link: http://hdl.handle.net/11104/0188494