Kernelized Constrained Gaussian Fields and Harmonic Functions for Semi-supervised Learning

Graph-based semi-supervised learning (SSL) methods are effective on many application domains. Despite such an effectiveness, many of these methods are transductive in nature, being uncapable to provide generalization for the entire sample space. In this paper, we generalize three existing graph-based transductive methods through kernel expansions on reproducing kernel Hilbert spaces. In addition, our methods can easily generate an inductive model in a parameter-free way, given a graph Laplacian constructed from both labeled and unlabeled examples and a label matrix. Through experiments on benchmark data sets, we show that the proposed methods are effective on inductive SSL tasks in comparison to manifold regularization methods.