Hyperplane Margin Classifiers on the Multinomial Manifold
Guy Lebanon - Carnegie Mellon University John Lafferty - Carnegie Mellon University
The assumptions behind linear classifiers for categorical data are examined and reformulated in the context of the multinomial manifold, the simplex of multinomial models furnished with the Riemannian structureinduced by the Fisher information. This leads to a new view of hyperplaneclassifiers which, together with a generalized margin concept, shows how toadapt existing margin-based hyperplane models to multinomial geometry. Experiments show the new classification framework to be effective for textclassification, where the categorical structure of the data is modelednaturally within the multinomial family.