Supervised Nonlinear Factorizations Excel In Semi-supervised Regression

Semi-supervised learning is an eminent domain of machine learning focusing on real-life problems where the labeled data instances are scarce. This paper innovatively extends existing factorization models into a supervised nonlinear factorization. The current state of the art methods for semi-supervised regression are based on supervised manifold regularization. In contrast, the latent data constructed by the proposed method jointly reconstructs both the observed predictors and target variables via generative-style nonlinear functions. Dual-form solutions of the nonlinear functions and a stochastic gradient descent technique which learns the low dimensionality data are introduced. The validity of our method is demonstrated in a series of experiments against five state-ofart baselines, clearly improving the prediction accuracy in eleven real-life data sets.