A Novel Feature Extraction Method - alpha-Based Supervised Orthogonal Projection Reduction by Affinity

摘要

In this paper, a novel feature extraction approach alpha-based supervised orthogonal projection reduction by affinity is proposed by introducing the idea of SLLE into the traditional method of OPRA. By adding an additional parameter a to control the degree of supervision, the proposed method can acquire some compromise between purely supervised OPRA and unsupervised OPRA and does not only keep the reservation of some flow-shaped structure during high-dimensional to low-dimensional mapping, but also gets better orthogonal projection. Experimental results based on both synthetic data and real data (human face recognition) show that the proposed method is more effective than either purely supervised OPRA or unsupervised OPRA and some other traditional feature extraction methods.