Some experimental results on learning probabilistic and possibilistic networks with different evaluation measures

Abstract

Abstract. A large part of recent research on probabilistic and possibilistic inference networks has been devoted to learning them from data. In this paper we discuss two search methods and several evaluation measures usable for this task. We consider a scheme for evaluating induced networks and present experimental results obtained from an application of INES (Induction of NEtwork Structures), a prototype implementation of the described methods and measures.

Data Mining with Graphical Models

“…as a measure of the quality of a given graphical model G [Borgelt and Kruse 1997b]. Obviously, this measure is very similar to the likelihood measure that may be used in the probabilistic case (cf.…”

“…Since we are trying to minimize the value of the measure, it seems natural to choose pessimistically the maximum as the worst possible case. This choice has the additional advantage that it can be computed efficiently by simply propagating the evidence contained in an imprecise tuple in the given graphical model [Borgelt and Kruse 1997b], whereas other aggregates suffer from the fact that we have to compute explicitly the degree of possibility of the compatible precise tuples, the number of which can be very large.…”

“…Following an introduction to the principles of global structure learning, which is intended to provide-like the greater part of chapter 3-an intuitive background, I discuss some well-known and suggest some new evaluation measures for learning probabilistic as well as possibilistic networks [Borgelt and Kruse 1997a, Borgelt and Kruse 1999b. Furthermore, I review some well-known and suggest some new search methods.…”

“…Therefore recent research has concentrated on learning graphical models from databases of sample cases (cf. [Herskovits and Cooper 1990, Cooper and Herskovits 1992, Buntine 1994, Jordan 1998] for learning probabilistic networks and [Gebhardt and Kruse 1995, Gebhardt and Kruse 1996b, Gebhardt and Kruse 1996c, Borgelt and Kruse 1997a, Borgelt and Kruse 1997b, Borgelt and Gebhardt 1997 for learning possibilistic networks). Thus graphical models entered the realm of data mining methods.…”

“…the average or the maximum, of the degrees of possibility of all precise tuples that are at least as specific as 5 an imprecise tuple [Borgelt and Kruse 1997b]. Since we are trying to minimize the value of the measure, it seems natural to choose pessimistically the maximum as the worst possible case.…”