Bayesian classification by iterated weighting

It is shown that the posterior probabilities of given classes can be learned in two steps: first, the likelihood functions are estimated by parametric optimization, then the class probabilities, or weights, are estimated by iterated averaging. The iterative procedure is proved to converge for the two-class problem. The separate calculation of the likelihoods and the weights makes the proposed approach "modular", allowing for the use of any density estimation method. It represents structural and computational simplification of the expectation maximization approach, which iterates on the density parameters and the weights simultaneously. Applied to the classification of medical and financial data, the average performance of the proposed iterated weighting method is found to be, in the first case, superior, and, in the second, very similar, to that of the expectation maximization method. The advantage of the proposed method is in its relative simplicity. (C) 1999 Elsevier Science B.V. All rights reserved.