Logistic regression is generally pretty good at being calibrated. This paper uses probit regression, but the basics are the same. The choice of link function affects whether there is an online update, not the calibration issue.
http://research.microsoft.com/apps/pubs/default.aspx?id=122779 My comment had to do with the assumptions of independence of features. If you have independent features, or if you include the necessary interaction features and regularize correctly, you should get pretty decent calibration. On Mon, Aug 5, 2013 at 8:36 AM, Jerome Williams <[email protected]>wrote: > I read the following thread > > http://markmail.org/message/4tgpbojg66hbml4i > > > mentioning that Bayesian classifier and logistics regression do not > produce calibrated probability. > > So my question - is there any algorithm that produce calibrated > probability (e.g. between 0 and 1)? > > Thanks. >
