How to compare MHMMs with different number of parameters?

[helske]

How to compare MHMMs with different number of parameters?

[satuhelske]

Model selection of MHMMs is an open issue. Here is one (imperfect) option: https://lacosa.lives-nccr.ch/sites/lacosa.lives-nccr.ch/files/proc-lacosa2-helskehelskeeerola_paper_24.pdf

In short, this approach starts with finding preliminary clusters using sequence analysis and cluster analysis, followed by estimating separate HMMs for each cluster in a few clustering options. In other words, the number of states in each preliminary cluster is determined separately. Finally, the parameters of these separate HMMs are used as starting values in the estimation of MHMMs, keeping numbers of hidden states fixed in each cluster. The best number of clusters is determined from combining the MHMMs.

Please read the paper for more detailed information to determine whether something like this would suit your needs.

[vrodriguezf]

Thanks! What section of the paper talks about this?

[helske]

Closed as old and slightly off-topic (this is more about general MHMM modelling instead of about the use of seqHMM).

In the paper linked above, section 6.2.2 discusses this issue shortly.