Two-Dimensional Neighborhood Structure Preserving Projection

This paper presents a novel manifold learning method, called two-dimensional neighborhood structure preserving projection (2DNSPP) for dimensionality reduction. 2DNSPP employs two adjacency graphs, namely diversity graph and similarity graph, with a vertex set and two affinity matrices. The affinity matrix of diversity graph characterizes the spatial diversity structure among nearby data, while affinity matrix of similarity graph characterizes the spatial similarity structure among nearby data. A concise feature extraction is then raised via combining the diversity and similarity among nearby data. Experiment results on the UMIST database indicate the efficiency of the proposed method.