Study of the classification in the subspace of the asymmetric principle component analysis

Abstract

My dissertation mainly studies the process of principal component analysis method

which widely used for pattern classification. Besides, it analyses the problems of

principal component analysis method when the training data are unbalance. A new

method called asymmetric principal component analysis (APCA) is used to remove

the less reliable dimensions to help boost the classification accuracy.

When dealing with a two-class classification problem, the discriminant analysis in

the APCA subspace is used to adjust the eigenvalues so that we can produce more

discriminative and reliable features for the asymmetric classes training data. We have

compared this approach with other approaches. The experimental results show the

highest accuracy among other approaches. We further find out that the optimal

weight factor of different type of training classes have some relationship with the

distribution of the training data.