Title: Sparse finite mixture-of-experts modeling
Authors: Gregor Zens - Bocconi University (Italy) [presenting]
Abstract: A sparse Bayesian mixture-of-experts model is developed where the number of components is estimated endogenously. Bayesian inference is carried out through the implementation of a Gibbs sampler. By constructing a suitable prior density for the component weights, we allow the sampler to apply shrinkage to the component weights when necessary. Moreover, additional information is allowed to enter the model via a set of independent covariates. This is achieved by combining the clustering information of the mixture-of-experts linking function and the idea of a deliberately overfitting mixture model. An identified model is obtained by relabeling the MCMC output in the point process representation of the draws. The model is evaluated in a simulation setup with artificial data. In a real data application, the model is used to find homogenous clusters of women in Mozambique based on their information sources on HIV.