Local Marginal Projection and Its Applications

Based on UDP and MFA, we propose a new un-supervised feature extraction algorithm, LMP (Local Marginal Projection), which is built on local quality. It measures the non-local quantities by the nearest sample between two locals. The goal of LMP is to find a projection that can maximize the distance of the sample in the same local and in different locals, in which case, the data can be projected into low-dimension easily. Besides, this projection could deal with the nonlinear and high-dimensional problem. The experiment on ORL and Yale face Database shows that LMP algorithm can describe the high-dimensional data and can embed the nonlinear data Swiss-Hole into low-dimension space with a reasonable visual effectively.