As we know mixture models are important tools in density estimation and in general in statistical machine learning. I have always used nonparametric Bayesian mixture models to avoid the problem of finding the optimal number of mixture ecomponents. Considering the fact that mixture models with nonparametric priors are quite old and well known, why is that I still see papers that use finite mixture models and then use a model selection method to find the best model. Do these methods offer other advantages over nonparametric models?

In particular, could anyone please let me know if the Minimum Message Length method have an advantage over using a nonparametric prior such as Dirichlet process?

  • $\begingroup$ Two simple reasons: (1) FMM's are more computationally efficient; (2) when estimating Bayesian mixtures with MCMC you have issues with label switching, with EM for FMM's you don't. $\endgroup$ – Tim Aug 1 '16 at 7:52
  • $\begingroup$ @Tim Is label switching an issue when the main purpose is clustering? $\endgroup$ – Adham Aug 1 '16 at 10:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.