A0860
Title: Semi-supervised learning using elliptical distributions with unknown density generators
Authors: Chin-Tsang Chiang - National Taiwan University (Taiwan) [presenting]
Abstract: A more general elliptical distribution model is proposed for the classification of the groups with both labeled and unlabeled data. Different from existing multivariate normal and t distribution models in semi-supervised learning, the forms of the density generators are left unspecified. By incorporating the information of unlabeled data into training, a pseudo maximum likelihood method is developed to estimate the finite-dimensional model parameters. An efficient computational procedure is further presented to perform the maximization of the pseudo-likelihood function. Especially, the proposed estimators of the posterior group probabilities are useful for constructing a prediction rule. In addition, our estimators are shown to be asymptotically more efficient than the corresponding ones using only labeled data. Simulations and applications to empirical data are also used to illustrate the methodology.
