Using incremental general regression neural network for learning mixture models from incomplete data

Finite mixture models (FMM) is a well-known pattern recognition method, in which parameters are commonly determined from complete data using the Expectation Maximization (EM) algorithm. In this paper, a new algorithm is proposed to determine FMM parameters from incomplete data. Compared with a modified EM algorithm that is proposed earlier the proposed algorithm has better performance than the modified EM algorithm when the dimensions containing missing values are at least moderately correlated with some of the complete dimensions.