A note on variational Bayesian factor analysis

Neural Netw. 2009 Sep;22(7):988-97. doi: 10.1016/j.neunet.2008.11.002. Epub 2008 Nov 25.


Existing works on variational bayesian (VB) treatment for factor analysis (FA) model such as [Ghahramani, Z., & Beal, M. (2000). Variational inference for Bayesian mixture of factor analysers. In Advances in neural information proceeding systems. Cambridge, MA: MIT Press; Nielsen, F. B. (2004). Variational approach to factor analysis and related models. Master's thesis, The Institute of Informatics and Mathematical Modelling, Technical University of Denmark.] are found theoretically and empirically to suffer two problems: (1) penalize the model more heavily than BIC and (2) perform unsatisfactorily in low noise cases as redundant factors can not be effectively suppressed. A novel VB treatment is proposed in this paper to resolve the two problems and a simulation study is conducted to testify its improved performance over existing treatments.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Algorithms*
  • Bayes Theorem*
  • Computer Simulation*
  • Humans
  • Neural Networks, Computer*
  • Pattern Recognition, Automated