An efficient kernel-based feature extraction using a pull-push method

Abstract Supervised dimensionality reduction methods using nonlinear mappings for pattern recognition tasks are more appropriate for nonlinearly distributed data. Generally, for most algorithms, these samples (called hard samples) located at the edge or in other classes influence the performance of the proposed methods. In this study, the large margin nearest neighbor (LMNN) and weighted local modularity (WLM) in the complex network are introduced to deal with these hard samples to push and pull them rapidly toward the center of the class, and the samples with the same label shrink, as a whole, into the center of the class. A novel feature extraction method named KWLM-LMNN, which uses the kernel trick and combining WLM with LMNN, is proposed. Comparative experiments with state-of-the-art feature extraction methods demonstrate the effectiveness of the proposed method.