A Semi-Supervised Learning Algorithm Based on a Hierarchical GMM

A semi-supervised learning algorithm based on a hierarchical GMM is proposed. The learning samples in semi-supervised learning are a hybrid of labeled and unlabeled samples. If GMM is employed to fit the distribution of labeled samples in each class and a hierarchical GMM whose Gaussian number is the class number is employed to fit the distribution of the whole learning samples (including labeled and unlabeled samples), then a semi-supervised learning problem based on a hierarchical GMM has emerged. Based on EM algorithm, by learning the labeled samples of each class, a GMM is obtained first. Then by taking the parameters of the obtained GMM and frequencies of labeled samples as initials, a semi-supervised learning algorithm based on a hierarchical GMM is presented. Printed numerals in a bank check are tested in the experiments and the results shows the good effects of the proposed algorithm.