Leave-One-Out Cross-Validation Based Model Selection for Manifold Regularization.

Classified labels are expensive by virtue of the utilization of field knowledge while the unlabeled data contains significant information, which can not be explored by supervised learning. The Manifold Regularization (MR) based semi-supervised learning (SSL) could explores information from both labeled and unlabeled data. Moreover, the model selection of MR seriously affects its predictive performance due to the inherent additional geometry regularizer of SSL. In this paper, a leave-one-out cross-validation based PRESS criterion is first presented for model selection of MR to choose appropriate regularization coefficients and kernel parameters. The Manifold regularization and model selection algorithm are employed to a real-life benchmark dataset. The proposed approach, leveraged by effectively exploiting the embedded intrinsic geometric manifolds, outperforms the original MR and supervised learning approaches. [ABSTRACT FROM AUTHOR]