A Supervised Manifold Learning Method

Zuojin Li, Weiren Shi, Xin Shi, Zhi Zhong

The Locally Linear Embedding (LLE) algorithm is an unsupervised nonlinear dimensionality-reduction method, which reports a low recognition rate in classification because it gives no consideration to the label information of sample distribution. In this paper, a classification method of supervised LLE (SLLE) based on Linear Discriminant Analysis (LDA) is proposed. First, samples are classified according to their label values, and low dimensional features of intraclass data are expressed through LLE manifold learning. Then, the base vectors in Fisher subspace of the low dimensional features are generated through LDA learning. This method increases inter-class variation, and decreases the intra-class variation when samples are projected to the Fisher subspace. Hence, the samples of different labels can be recognized, and the recognition rate and robustness of the LLE learning are improved. Experiments on handwritten digit recognition show that the proposed method is featuring high recognition rate.