TSFS: A Novel Algorithm for Single View Co-training

Co-training has been validated to be effective in various applications. However, it is a challenging task to apply co-training on the data without two independent and "good enough" views. In this paper, we propose a novel subspace feature set splitting algorithm, called Two-view Subspace Feature Splitting (TSFS), to make co-training better usable on single view data. We first project both labeled and unlabeled data into a lower dimensional subspace through Singular Value Decomposition (SVD), in which all features of data are orthogonal to each other. And then a greedy two-view feature selection strategy is proposed for feature set splitting. We introduce the energy function of each view to guarantee the quality of each split feature set. Experimental results well validated the effectiveness of TSFS in contrast to several recent studies on single view co-training.