Sparse Mixture-Of-Experts Models

Description

Selecting the appropriate number of clusters is a long-standing problem in mixture modeling and model-based clustering in general. A number of successful approaches propose to treat the number of clusters as a random quantity to be estimated. However, such methodology is not available for mixture-of-experts models, a model class where covariates are used to inform cluster membership. We aim to fill this gap and develop a flexible Bayesian mixture framework that combines covariate-dependent mixture weights and endogenous estimation of the number of mixture clusters. We show that model selection procedures based on overfitting finite mixtures can be extended to the class of mixture-of-experts models and derive the implied prior distributions. Finally, we outline an accurate Markov chain Monte Carlo sampling scheme for efficient posterior simulation. The utility of the framework is illustrated using simulated data and a real world example.

Period	20 Apr 2022 → 22 Apr 2022
Event title	Austrian and Slovenian Statistical Days 2022
Event type	Unknown
Degree of Recognition	National