|
Hi, I am looking for something like an Indian Buffet Process, only the other way around. That is, I'd like to go from binary features to continues latent features. Non-parametricity (if there is such a word) is not critical, meaning the number of latent features can be constant. I know of mixtures of Bernoullis, but with MoB it seems that the model assumes that each data point is explained by exactly one component, and I'd like the points to be explained by more than one component (think Chinese restaurant vs. Indian Buffet). I also know of RBMs. Is there anything else? Please point me to concepts and/or software implementations. |
|
You could look at Hierarchical models, if non-parametric is not a requirement, you could do hierarchical mixtures of whatever you want, that way each feature can be part of different features. I do not know if I'm following you but I do not see the need to describe your problem in terms of CRP and IBP. Remember that both are just infinite generalizations to simpler models, in the case of the IBP you have a Factor Analysis model. You could do a hierarchical mixture of Bernoulli, where a Dirichlet distribution controls them. |
