Sequence Classification in the Jensen-Shannon Embedding

This paper presents a novel approach to the supervised classification of structured objects such as sequences, trees and graphs, when the input instances are characterized by probability distributions. Distances between distributions are computed via the JensenShannon (JS) divergence, which offers several advantages over the L2 distance or the Kullback-Leibler divergence. The JS divergence induces an embedding of the distributions into a real Hilbert space. A general approach is proposed here to derive a positive definite kernel from any conditionally negative definite (CND) distance, the JS divergence being a particular case of interest. We show how to compute the dot product in the embedding induced by a CND distance, based solely on the distance matrix between training points, and we detail how new points can be added to this embedding. The JS kernel is applied to sequence classification problems. Two kinds of empirical distributions are considered: (i) the N -gram distributions and (ii) the distributions of the First Passage Times (FPT) between occurrences of substrings. Experimental results on DNA splicing junction detection and protein function prediction illustrate that . . . Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute.