Topic models can be applied to any dataset that has group structure.
FiniteTopicModel supports VB, soVB, moVB
HDPTopicModel supports VB, soVB, and moVB. * with birth/merge/delete moves for moVB
FiniteTopicModel: stuff here
HDPTopicModel: more stuff here
There are two types of mixture model supported. Both define the model in terms of a global parameter vector \(\beta\), where \(\beta_k\) gives the probability of topic k, and local assignments \(z\), where \(z_n\) indicates which state {1, 2, 3, … K} is assigned to data item n.
The FiniteMixtureModel has a generative process:
while the DPMixtureModel has generative process:
If we let K grow to infinity, these two models converge if \(\gamma = \gamma_0 /K\).