A0345
Title: Recent developments in variational inference algorithms for restricted latent class models
Authors: Kazuhiro Yamaguchi - University of Tsukuba (Japan) [presenting]
Abstract: As diagnostic classification models, restricted latent class (RLC) models have been employed in the social sciences, especially in psychology or educational research. The RLC is a special case of a general latent class model or mixture model in which the latent variables are categorical, and the latent classes are used to classify individuals into understandable sub-populations. Information and communication technology provides a wealth of rich information sources for latent classes and enables extended RCL models. However, an increase in data sources can make it difficult to estimate the model parameter of complex RCL models. The Variational Bayesian (VB) inference method, employed for complex machine learning models, is a good choice for RLC models. It is a deterministic posterior approximation method that works faster than the Markov chain Monte Carlo method. The key to deriving the VB estimation algorithm for an RLC model is introducing an auxiliary variable to represent equality constraints on the general latent class models. These constraint variables provide tractable mean-field variational inference for the RCL models. Furthermore, the VB method for extended RCL models (such as hidden Markov RCL or two-level RCLs models) can be derived based on this formulation. Recent developments in VB inference for various types of RLC models were reviewed and discussed.
