The kernelHMM: Learning kernel combinations in structured output domains

Abstract

We present a model for learning convex kernel combinations in classification problems with structured output domains. The main ingredient is a hidden Markov model which forms a layered directed graph. Each individual layer represents a multilabel version of nonlinear kernel discriminant analysis for estimating the emission probabilities. These kernel learning machines are equipped with a mechanism for finding convex combinations of kernel matrices. The resulting kernelHMM can handle multiple partial paths through the label hierarchy in a consistent way. Efficient approximation algorithms allow us to train the model to large-scale learning problems. Applied to the problem of document categorization, the method exhibits excellent predictive performance. © Springer-Verlag Berlin Heidelberg 2007.