Laplacian regularized co-training

Co-training is one promising paradigm of semi-supervised learning and has drawn considerable attentions and interests in recent years. It usually works in an iterative manner on two disjoint view features, in which two classifiers are trained on the different views and teach each other by adding the predictions of unlabeled data to the training set of the other view. However, the classifier performs not well with small number of labeled examples especially in the first rounds of interation. In this paper, we present Laplacian regularized co-training(LapCo) to address the above problem in standard co-training. During the training process, LapCo employs Laplacian regularization into the classifier to significantly boost the classification performance. The experiments on three popular UCI repository datasets are conducted and show that the proposed LapCo outperforms the traditional co-training method.