Projective Label Propagation by Label Embedding

In this paper, we propose a projective label propagation (ProjLP) framework by label embedding that can gain the more discriminating “deep” labels of points in an transductive fashion to enhance representation. To show the deep property of the embedded “deep” labels over the “shallow” ones that usually have unfavorable mixed signs delivered by existing transductive models, the auxiliary multilayer network architecture of our ProjLP is illustrated. The deep architecture has three layers (i.e., input layer, hidden layer, and output layer). For semi-supervised classification, ProjLP delivers the deep labels of data via two-layer label propagation (i.e., hidden and output layer) on the network at each iteration. In hidden layer, ProjLP delivers the “shallow” soft labels

F

of points in the original input space. Then, ProjLP embeds

F

onto a subspace spanned by a robust projection to obtain the deep soft labels in output layer. Finally, the most discriminating deep labels are obtained for enhancing performance. The method of achieving the deep labels of outside points is also elaborated. Simulations on several artificial and UCI datasets demonstrate the validity of our model, compared with other state-of-the-arts.