9.3 The difficulty of Bayesian inference for clustering
Two problems make it pretty much impossible to perform full Bayesian inference for clustering models, the lack of parameter identifiability and the extreme multimodality of the posteriors. There is additional discussion related to the non-identifiability due to label switching in the label switching section.
Non-identifiability
Cluster assignments are not identified—permuting the cluster mean
vectors mu leads to a model with identical likelihoods. For
instance, permuting the first two indexes in mu and the first
two indexes in each soft_z[n] leads to an identical likelihood
(and prior).
The lack of identifiability means that the cluster parameters cannot be compared across multiple Markov chains. In fact, the only parameter in soft \(K\)-means is not identified, leading to problems in monitoring convergence. Clusters can even fail to be identified within a single chain, with indices swapping if the chain is long enough or the data are not cleanly separated.
Multimodality
The other problem with clustering models is that their posteriors are highly multimodal. One form of multimodality is the non-identifiability leading to index swapping. But even without the index problems the posteriors are highly multimodal.
Bayesian inference fails in cases of high multimodality because there is no way to visit all of the modes in the posterior in appropriate proportions and thus no way to evaluate integrals involved in posterior predictive inference.
In light of these two problems, the advice often given in fitting clustering models is to try many different initializations and select the sample with the highest overall probability. It is also popular to use optimization-based point estimators such as expectation maximization or variational Bayes, which can be much more efficient than sampling-based approaches.