|
Abstract:
|
We propose a prior for a class of infinite latent feature models, in which each latent feature is assumed to be expressed as a product of two latent variables, a binary latent variable for its representation and a latent score variable for its value. The prior for a collection of the binary latent variables is constructed by assuming binomial distributions with probabilities of success stochastically decreasing in the index of the latent variables, while that for the latent scores is assigned a Dirichlet process prior. Upon marginalizing out the latent features, a mixture model is induced, allowing us to make a variety of inferences. Noting that the proposed approach results in a model having a sparse representation with a few dominant features, we consider a truncation approximation for the practical implementation. An efficient MCMC algorithm is developed for posterior computation.
|
