Bayesian model averaging is not model combination

Thomas P. Minka
MIT Media Lab note (7/6/00)

In a recent paper, Domingos (2000) compares Bayesian model averaging (BMA) to other model combination methods on some benchmark data sets, is surprised that BMA performs worst, and suggests that BMA may be flawed. These results are actually not surprising, especially in light of an earlier paper by Domingos (1997) where it was shown that model combination works by enriching the space of hypotheses, not by approximating a Bayesian model average. And the only flaw with BMA is the belief that it is an algorithm for model combination, when it is not.

Postscript . PDF

Last modified: Fri Dec 10 14:31:21 GMT 2004