Behavioral Learning for Data Adjacent Graph Construction in Semi-supervised Learning

Laplacian support vector machine could utilize the unlabeled samples for semi-supervised learning by applying the manifold regularization term. But the data adjacent graph in the manifold regularization term couldn't take advantage of the label information and the empirical setting of heat kernel parameter would also degrade the learning performance. Inspired by human behavioral learning theory, a novel semi-supervised learning with local behavioral similarity was proposed in this paper to solve those problems. In detail, the new edge weight with label information was introduced and the local distribution parameter considering the underlying probability distribution in the neighborhood of a point was applied. Extensive experiments on public data sets show the good performance and validity of the new algorithm.