A Study of Distance Functions in FastMapSVM for Classifying Seismograms

FastMapSVM is a recently developed Machine Learning framework that combines the complementary strengths of FastMap and SVMs for classification tasks. It is particularly useful when it is easier to measure the dissimilarity between pairs of objects in the domain via a well-defined distance function on them than it is to identify and reason about complex characteristic features of individual objects. The success of FastMapSVM has also been recently demonstrated in the Earthquake Science domain, where the objects are seismograms that need to be classified as earthquake signals or noise signals. In this paper, we first define various distance functions on seismograms. We then study the effects of these different distance functions on the performance characteristics of FastMapSVM. We also evaluate the different distance functions on their ability to provide perspicuous visualizations of the seismograms, their spread, and the classification boundaries between them.