Kernel Feature Maps from Arbitrary Distance Metrics

Abstract

Listen

Translate article

AbstractThe approximation of kernel functions using explicit feature maps gained a lot of attention in recent years due to the tremendous speed up in training and learning time of kernel-based algorithms, making them applicable to very large-scale problems. For example, approximations based on random Fourier features are an efficient way to create feature maps for a certain class of scale invariant kernel functions. However, there are still many kernels for which there exists no algorithm to derive such maps. In this work we propose an efficient method to create approximate feature maps from an arbitrary distance metric using pseudo line projections called Distance-Based Feature Map (DBFM). We show that our approximation does not depend on the input dataset size or the dimension of the input space. We experimentally evaluate our approach on two real datasets using two metric and one non-metric distance function.KeywordsSupport Vector MachineReproduce Kernel Hilbert SpaceRandom ProjectionScene CategoryHellinger DistanceThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.