Trace ratio-based dimensionality reduction for discriminative analytics of multiple datasets

Abstract:

Principal component analysis is widely applied in dimensionality reduction and feature extraction，especially in tackling single high-dimensional dataset. However，traditional principal component analysis faces challenge when it comes to analyzing multiple datasets jointly. This paper introduces a novel approach named trace ratio principal component analysis，which can discover low-dimensional structure unique to the target data relative to others. Furthermore，trace ratio principal component analysis and its variants can be solved by efficient iterative algorithm. Numerical experiments show the efficiency of the method.