Data Structure Based Discriminant Score for Feature Selection

Selecting features from hyperspectral data under unsupervised mode is a hard work, owing to the absence of labeled data. However, most of current unsupervised feature selection algorithms ignore the fact that real data has the distribution of manifold structure which is embedded into original high dimensional space. In order to solve this problem, an unsupervised feature selection method based on the data structure, called Data structure based Discriminant Score (DDS) is presented in this paper. The proposed algorithm is a linear approximation of multi-manifolds based process which considering local and non-local quantities simultaneously. It evaluates candidate features by calculating their power of maximizing the non-local, and in the same time, minimizing the local scatter. The property enables DDS more effective than some other feature selection methods. Experiments on a benchmark hyperspectral data set demonstrate the efficiency of our algorithm.