Local patch effects on LPP and DLA in dimensionality reduction

In order to preserve the local structure of the input data, many dimensionality reduction algorithms are involved in the definition of certain locality in the feature space. Thus local patch selection will directly affect the performance of these algorithms. In this paper, we propose a new method of local patch selection for dimensionality reduction. Specially, we incorporate our method into locality-preserving projection (LPP) and discriminative locality alignment (DLA), which are two typical algorithms without and with using label information in local patch selection. We do the experiments for scene classification with the Object Bank of 2124 original dimensions. We provide comprehensive evaluation of patch selection methods and extensive analysis on the performance of different methods.