Boosting and Naive Bayesian Learning

Although so-called “naive” Bayesian classification makes the unrealistic assumption that the values of the attributes of an example are independent given the class of the example, this learning method is remarkably successful in practice, and no uniformly better learning method is known. Boosting is a general method of combining multiple classifiers due to Yoav Freund and Rob Schapire. This paper shows that boosting applied to naive Bayesian classifiers yields combination classifiers that are representationally equivalent to standard feedforward multilayer perceptrons. (An ancillary result is that naive Bayesian classification is a nonparametric, nonlinear generalization of logistic regression.) As a training algorithm, boosted naive Bayesian learning is quite different from backpropagation, and has definite advantages. Boosting requires only linear time and constant space, and hidden nodes are learned incrementally, starting with the most important. On the realworld datasets on which the method has been tried so far, generalization performance is as good as or better than the best published result using any other learning method. Unlike all other standard learning algorithms, naive Bayesian learning, with and without boosting, can be done in logarithmic time with a linear number of parallel computing units. Accordingly, these learning methods are highly plausible computationally as models of animal learning. Other arguments suggest that they are plausible behaviorally also.