Statistical pattern recognition: Locality preserving embeddings and ensemble of rules

In this work, we contribute to develop 4 new techniques for dimensionality reduction based on the above mentioned methods.;Firstly, we develop the 2D Laplacianfaces by intergrating the two techniques, locality preserving and image-based projection. The training time and memory complexity is reduced from O(m2 × n2) to only O( m × n), where m and n are the number of rows and columns of the sample image. Secondly, we develop a technique, unsupervised discriminant projection (UDP). In addition to the local information, we also consider the global information in formulating the optimizing criterion. Thirdly, we propose Mutual Neighborhood based Discrimiant Projection (MNDP). We construct the mutual neighborhoods to highlight those samples that are on the boundary and most likely contribute to the prediction errors. Finally, we present adaptive CS4 (ACS4) for feature selection and prediction. ACS4 can identify the most discriminant features and grow a committee of trees for prediction. We evaluate ACS4 on the biology database for DNA methylation analysis.