Small sets of random Fourier features by kernelized Matrix LVQ
Frank-Michael Schleif · 2017
Kernel based learning is very popular in machine learning but often quite costly with at least quadratic runtime complexity. Random Fourier features and related techniques have been proposed to provide an explicit kernel expansion such that standard techniques with low runtime and memory complexity can be used. This strategy leads to rather high dimensional datasets which is a drawback in many cases. Here, we combine a recently proposed unsupervised selection strategy for random Fourier features [1] with the very efficient supervised relevance learning given by Matrix LVQ. The suggested technique provides reasonable small but very discriminative features sets1.