Semi-supervised learning scheme using Dirichlet process EM-algorithm

摘要

Learning with a dataset that contains both labeled data and unlabeled data is often called a semi-supervised learning problem. In the last decade, the semi-supervised learning problem has become an important research problem in many fields. This article presents a novel semi-supervised learning scheme using a Bayesian Maximum A Posteriori Expectation Maximization (MAP-EM) algorithm with a Dirichlet process prior (stick-breaking representation). The proposed scheme enables us to estimate a mixture model under an unknown number of components and provides a simpler implementation than other implementations such as Markov Chain Monte Carlo (MCMC) implementations. Several examples of a Gaussian mixture are examined to validate the proposed scheme.