"Learning of Highly-Filtered Data Manifold Using Spectral Methods"
O. Roderick and I. Safro
Preprint Version: [pdf]
We propose a scheme for improving existing singular value decomposition-based tools for recovering and predicting decisions. Our main contribution is an investigation of advantages of using a functional, rather than popular linear approximation of the response of an unknown, complex model. A significant attractive feature of the method is the demonstrated ability to make predictions based on a highly filtered data set.