Semi-Supervised Spectral Mapping for Enhancing Separation between Classes

We present a spectral mapping technique for semisupervised pattern classification. Importance scores of features are firstly evaluated with a semi-supervised feature selection algorithm by Zhao et al. Training data are then embedded into a low-dimensional space with a spectral mapping derived from the selected and weighted feature vectors with which test data are classified by the nearest neighbor rule. The performance of the proposed pattern classification algorithm is examined with synthetic and real datasets.