Feature Extraction Using Linear and Non-linear Subspace Techniques

This paper provides a new insight into unsupervised feature extraction techniques based on subspace models. In this work the subspace models are described exploiting the dual form of the basis vectors. In what concerns the kernel based model, a computationally less demanding model based on incomplete Cholesky decomposition is also introduced. An online benchmark data set allows the evaluation of the feature; extraction methods comparing the performance of two classifiers having as input the raw data and the new representations.