Hi all,
A widely used measure for the performance of algorithms that learn
Bayesian network from data is cross-entropy (Kullback & Leibler,
1951). It seems to me (maybe I'm wrong) that cross-entropy is used in
different ways by different researchers (for example Heckerman, Geiger
& Chickering, 1995 vs Lam & Bacchus, 1994). Cross-entropy seems to be
a general purpose measure that is used to quantify the distance
between two distributions (first: gold standard model - second:
induced model ), but the nature of these distributions can be
different for different researchers. What are the theoretical
implications of these differences ? Is there any paper comparing the
properties of various performance measures for Bayesian network
inducers ? Can anyone give me suggestions for readings on the
cross-entropy topic ?
Thanks to all.
Fabio Del Missier
====================================
Fabio Del Missier
PhD student
Department of Psychology, Univ. of Trieste.
Via S. Anastasio, 12, 34123, Trieste, Italy
tel: +39 040 676 2716
e-mail: [EMAIL PROTECTED]
====================================
