Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....of descriptions, rather than a single Classic concept; in other words, to learn a target concept c j d 1 d 2 : dn where each d i is a Classic2 description. Learning disjunctions of Classic concepts is somewhat analogous to the problem of inductive logic programming (ILP) Quinlan, 1990; Muggleton and Feng, 1992 ] In ILP the target concept is usually assumed to be a single Prolog predicate that is defined by a set of Prolog clauses; such a concept can often be viewed as a disjunction of the sets defined by each clause. Thus one natural approach to learning disjunctions of Classic descriptions is to ....

....viewed as a disjunction of the sets defined by each clause. Thus one natural approach to learning disjunctions of Classic descriptions is to adapt the techniques used in ILP to learn multi clause Prolog predicates. One well known ILP method for learning multiple clauses is the GOLEM algorithm [ Muggleton and Feng, 1992 ] which is also based on computing least common generalizations. The basic idea behind this algorithm is to use LCS to implement a specificto general greedy search for descriptions that cover many positive examples and no negative examples. In GOLEM, these descriptions are then further ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

....is true when jp 1 Gamma p 2 j 2. ffl near3 (p 1 ; p 2 ) is true when jp 1 Gamma p 2 j 3. ffl after(p 1 ; p 2 ) is true when p 2 p 1 . In experiments with FOIL we also used the relation succ(p 1 ; p 2 ) which is true when p 2 = p 1 1. This predicate was included since it is determinate [ Muggleton and Feng, 1992; Quinlan, 1991 ] and the FOIL algorithm includes special mechanisms that exploit determinism. ILP 95, Leuven 2.3 Evaluating performance for text categorization Text categorization learning problems tend to have certain common properties. There are usually a large number of features. It is ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

....1 Introduction Because learning arbitrary logic programs is difficult, efficient ILP systems have imposed a number of restrictions on the programs they learn. For example, LINUS learns only constrained clauses [ Dzeroski and Lavrac, 1991 ] GOLEM learns determinate clauses of bounded depth [ Muggleton and Feng, 1992 ] and CLINT learns clauses with few free variables [ Raedt and Bruynooghe, 1992 ] In addition, recent formal results [ Dzeroski et al. 1992; Cohen, 1993b; Cohen, 1993a ] have suggested a number of other potentially useful biases. However, there is as yet no clear understanding of how these ....

....number of arguments to the target predicate and the number of predicates declared, but exponential in the maximal arity j of a declared predicate. In the next few sections, we will use lazy macros to specify some less restrictive ILP biases. 3 Constant depth determinate clauses The GOLEM system [ Muggleton and Feng, 1992 ] learns ij determinate clauses. This bias or variations of it have subsequently been adopted by FOIL [ Quinlan, 1991 ] and LINUS [ Lavrac and Dzeroski, 1992 ] formally, it is known that a single ij determinate clause can be pac learned. 4 In this section we will describe how this bias can ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

....one clause programs. We will denote this language below a DetLP, where a is the bound on arity. Again, we will assume that the reader is familiar with this representation; however Appendix B contains a brief overview of the necessary background on logic programs. In previous work, Dzeroski, Muggleton and Russell [ 1992 ] showed that for any constant a, constant depth a DetLP programs are pac learnable. Later, Kietz [ 1993 ] showed that 2 DetLP programs of arbitrary depth are not paclearnable, and Cohen [ 1993 ] showed that 3 DetLP logdepth programs are not predictable. To date, however, the predictability of ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

....a propositional representation for text. These representations ignore such intuitively important properties as the order of words in a document. The first order learning techniques that have been recently developed in the machine learning community [ Quinlan, 1990; Quinlan and CameronJones, 1993; Muggleton and Feng, 1992 ] hold the potential for improved performance on this task, as they can (at least in principle) formulate classifiers that depend on word order. In this paper, we will experimentally evaluate the performance of FOIL, an off the shelf first order learning system, on text categorization problems. ....

....on matching regular expressions) do not allow retrieval of documents based on the absence of words. In such a context, the ability to learn a monotone classifier would be of value. Finally, we note that many first order learning systems learn pure Horn clauses, without negated literals (e.g. Muggleton and Feng, 1992 ] We thus re ran the experiments of Table 3, constraining FOIL6 to output monotone clauses. Table 4 summarizes the results. Contrary to our expectations, enforcing monotonicity did not improve generalization performance; rather, there is a general degradation of error rate, recall and ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

....candidate clause that can be proved to be true more than M distinct ways for any uncovered example is discarded. The intent is that by appropriately restricting M , a user can avoid expensive clauses, without having to commit to any particular syntactic restriction on clauses such as determinacy [ Muggleton and Feng, 1992 ] or locality [ Cohen, 1994b ] Lines M1 M2 of Table 2 show the performance of FLIPPER with modes and with M set to 100 and 500. Table 3: Comparison of FOIL and FLIPPER on artificial data time(FOIL) time(FLIPPER) error(FOIL) error(FLIPPER) k fi=0.0 0.1 0.2 0.25 0.3 fi=0.0 0.1 0.2 0.25 0.3 ....

Stephen Muggleton and Cao Feng. Efficient induction of logic programs. In Inductive Logic Programming. Academic Press, 1992.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC