A kernel principal component analysis algorithm based on sample selection according to pseudo-eigenvalue contribution

When kernel principal component analysis (KPCA) is applied for pattern classification problems such as face recognition, the more training samples are not leading to the easier way to get the useful principal components for classification. The reason is there are some samples containing lots of redundant information which is not conducive to classification and must be excluded from the training samples. Recent documents proposed an idea of electing points (samples) based on pseudo-eigenvalue method, deciding which point should remain in the training set according to their representation of the population. But the specific process and corresponding parameters of selecting points need to be further studied and improved. This article first discusses the specific process of selecting points and the rule of selecting parameters, and then applies the pseudo-eigenvalue method to KPCA to form a new PKPCA method. The experimental results based on ORL and YALE facial databases show that the proposed PKPCA algorithm has high efficiency and reliability.