Probabilistic Classifiers and the Concepts They Recognize

Proceedings of the Twentieth International Conference on Machine Learning

Volume

Proceedings of the Twentieth International Conference on Machine Learning

Contents

We investigate algebraic, logical, and geometric properties of concepts recognized by various classes of probabilistic classifiers. For this we introduce a natural hierarchy of probabilistic classifiers, the lowest level of which comprises the naive Bayesian classifiers. We show that the expressivity of classifiers on the different levels in the hierarchy is characterized algebraically by separability with polynomials of different degrees. A consequence of this result is that every linearly separable concept can be recognized by a naive Bayesian classifier. We contrast this result with negative results about the naive Bayesian classifier previously reported in the literature, and point out that these results only pertain to specific learning scenarios for naive Bayesian classifiers. We also present some logical and geometric characterizations of linearly separable concepts, thus providing additional intuitive insight into what concepts are recognizable by naive Bayesian classifiers.