Neighborhood Balance Embedding for Unsupervised Dimensionality Reduction
Mingming Sun, Chuancai Liu, Jingyu Yang · 2009
Various of manifold learning methods have been proposed to capture the intrinsic characteristic of nonlinear data. However, when confronting highly nonlinear data sets, existing algorithms may fail to discover the correct inner structure of data sets. In this paper, we proposed a new locality-based manifold learning method Neighborhood Balance Embedding. The proposed method share the same 'neighborhood preserving' property with other manifold learning methods, however, it describe the local structure in a different way, which makes each neighborhood like a s rigid balls, thus prevents the overlapping phenomenon which often happens when coping with highly nonlinear data. Experimental results on the data sets with high nonlinearity show good performances of the proposed method.