I have been debating between a model-based parametric clustering approach (e.g. HMMs), and a hierarchical Dirichlet/Pitman-Yor process for clustering sequential data. I understand the latter has been very popular in the recent past (no model selection, heavy tails, etc.), but I could not find any analysis of how it performs on really small datasets. I suspect that HMMs would be better, but I would really appreciate if someone could provide additional insight or point me to relevant texts on this topic.
